Markov Chain and Numerical Weather Prediction Models

Department of Physics and Technology
Short-term wind power prediction based on Markov

chain and numerical weather prediction models: A
case study of Fakken wind farm
Morten Jacobsen
EOM-3901 Masters Thesis in Energy, Climate and Environment
June 2014
ii
If you do tomorrow what you did today, you will get tomorrow what you got today.
Benjamin Franklin
iii
iv
Abstract
Rising energy demands and a growing focus on sustainable development have
made electricity production from wind energy an attractive alternative to
fossil fuels. However the natural variability of wind makes it challenging to
implement wind energy into the electrical grid. Accurate and reliable wind
power predictions are seen as a key element for an increased penetration of
wind energy.
This study presents a set of statistical power prediction models using
the concept of Markov chains, based on various input parameters, such as
wind speed, direction and power output. The models have been trained
and tested using numerical weather predictions and historical data obtained
from a meteorological station and wind turbine at Fakken wind farm in the
time period 2. May 2013 - 31. March 2014. Several of the models were found
to have lower NRMSE than the currently used persistent model (19.08 %),
with the best performing model having a NRMSE of 16.84 %. This 2.25 %
lower NRMSE corresponds to approximately 3.1 106 kWh of the anually
electricity production from Fakken wind farm.
A statistical analysis of Fakken wind farm showed the majority of winds
occurring from the straits between Arnya and Lenangsyra to the southeast and between Reinya and Lenangsyra to the south. Winds were also
commonly seen from southwest and to the northwest, while eastern and
northeastern winds were rarely observed. Westerly winds were found to be
much more tubulent than other directions, with a generally lower power
output observed. This is most likely due to the occurerence of mountain
waves for winds crossing the mountain range to the west.
vi
Acknowledgements
First of all I would like to thank my supervisor Yngve Birkelund for all the
time you have dedicated to help me with this thesis. The results would not
have been the same without your guidance and ideas. Not least for suggesting this thesis in the first place when things didnt quite work out as
expected with Dyry.
I would also like to thank Svein Erik Thyrhaug and Ronald Hardersen at
Troms Kraft for letting me use data from Fakken wind farm in my thesis and
for our meetings where you provided me with valuable information and feedback. I am very grateful for the time you have invested in me and this thesis.
A special thanks to Eirik Mikal Samuelsen at MET, for providing me with
the NWP data, teaching me about meteorology and answering my questions. I appreciate your friendliness and willingness to help me, even when
you were busy.
To my beloved Maria, always being there for me, supporting me and sticking
up with my many late hours. I love you very much.
To all the awesome people at Norutbrakka and especially Team Double
Bounce Scattering. This has been 5 incredible years I will never forget.
Last but not least, a big thanks to the universe for skipping spring in Troms
this semester and delaying the summer by several weeks. Nice and sunny
weather would have made this journey unbearable.
Morten Jacobsen
Troms, June 2014
vii
viii
Contents
List of figures
xi
List of tables
xvii
1 Introduction
1.1 Short-term wind power
1.2 Former research . . . .
1.3 Purpose of the study .
1.4 Structure of the study
prediction
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
2 Theory
2.1 Source of wind energy . . . . . . . .
2.2 Global wind patterns . . . . . . . . .
2.3 Vertical wind speed profile . . . . . .
2.4 Distribution of wind . . . . . . . . .
2.5 Wind power production . . . . . . .
2.6 Topographic features . . . . . . . . .
2.7 Numerical weather prediction model
2.8 General statistics . . . . . . . . . . .
2.9 Markov chains . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
5
5
.
.
.
.
.
.
.
.
.
7
7
8
10
13
16
27
28
28
35
3 Methods
39
3.1 Site and time . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Power prediction models . . . . . . . . . . . . . . . . . . . . . 46
4 Results
4.1 Rose diagrams .
4.2 Power curve . . .
4.3 Power, speed and
4.4 Markov model . .
. . . . . . . . . . .
. . . . . . . . . . .
direction diagrams
. . . . . . . . . . .
ix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
51
51
52
56
61
5 Discussion
5.1 Wind roses . . . . . . . . . . . . . .
5.2 Power curve . . . . . . . . . . . . . .
5.3 Power, speed and direction diagrams
5.4 Markov chain models . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
69
69
70
70
71
6 Conclusions
73
6.1 Further research . . . . . . . . . . . . . . . . . . . . . . . . . 74
Bibliography
75
Abbreviations
ABL
Atmospheric boundary layer
HAWT Horizontal axis wind turbine

MC
Markov chain
NRMSE Normalized root mean square error

NWP Numerical weather prediction
PM
Persistent model
RMSE Root mean square error

rpm
Rotations per minute
STD
Standard deviation
xi
xii
List of Figures
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
2.14
3.1
3.2
3.3
3.4
A simple convection cell, with low pressure denoted by L and

high pressure by H. . . . . . . . . . . . . . . . . . . . . . . . .
The global wind circulation [Guido, 2008]. . . . . . . . . . . .
Vertical layers of the atmospheric buondary layer over flat
homogenous terrain. . . . . . . . . . . . . . . . . . . . . . . .
Weibull probability density function as a function of wind
speed for various values of b. . . . . . . . . . . . . . . . . . .
Airflow through a HAWT. . . . . . . . . . . . . . . . . . . . .
Air particles flowing along an airfoil [Babinsky, 2003]. . . . .
Streamlines flowing along an airfoil with increased angle of
attack [Babinsky, 2003]. . . . . . . . . . . . . . . . . . . . . .
Wind incident on a horizontal axis wind turbine (a), with a
section of a turbine blade at length r from the rotor center
given in (b). . . . . . . . . . . . . . . . . . . . . . . . . . . .
Design of a modern HAWT [Schubel and Crossley, 2012]. . .
Power curve showing the relation between output power and
wind speed for a Vestas V90 3.0MW HAWT. Values in graph
are taken from [Vestas Wind Systems A/S, 2013]. . . . . . . .
Vertically propagating mountain waves [Muller, 2014]. . . . .
Vertically decaying mountain waves [Muller, 2014]. . . . . . .
Relation between kartesian and polar coordinates [Jammalamadaka and Sengupta, 2001]. . . . . . . . . . . . . . . . . . .
Example of a wind rose . . . . . . . . . . . . . . . . . . . . .
Map of the location and surrounding area of Fakken wind
farm with WTG08 and meteorology pointed out. . . . . . .
WAA151 anemometer, where (1) is the cup wheel assembly,
(2) the sensor shaft and (3) the lower body [Vaisala Oyj, 2002a].
WAV151 wind vane, where (1) is the vane assembly, (2) the
sensor shaft and (3) the lower body [Vaisala Oyj, 2002b]. . .
Algorithmic overview of the Markov chain wind power prediction model. . . . . . . . . . . . . . . . . . . . . . . . . . .
xiii
8
9
11
14
20
21
21
24
25
26
29
29
33
34
40
42
44
50
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10
4.11
4.12
4.13
4.14
Rose diagram of in-situ measurements from the meteorology

station at Fakken wind farm, in the period 2. May 2013 - 31.
March 2014 (UTC-time). . . . . . . . . . . . . . . . . . . . .
Rose diagram of the forecast wind at Fakken wind farm, in
the period 2. May 2013 - 31. March 2014 (UTC-time). . . . .
Fakken wind farm (red rectangle) with the surrounding area
and directions shown. . . . . . . . . . . . . . . . . . . . . . .
Rose diagram of in-situ measurements from the meteorology
station at Fakken wind farm, 27. October 2013 (UTC-time).
Rose diagram of the forecast wind at Fakken wind farm ,27.
October 2013 (UTC-time). . . . . . . . . . . . . . . . . . . .
Theoretical power curve (black) with measured power output
from WTG08 as a function of measured wind speed (blue
circles). Time period is 2. May 2013 - 31. March 2014 (UTCtime). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Theoretical power curve (black) with measured power output
from WTG08 as a function of forecast wind speed (red circles)
.Time period is 2. May 2013 - 31. March 2014 (UTC-time). .
Case study of 27. Ocotber 2013 (UTC-time). Theoretical power curve (black) with measured power output from
WTG08 as a function of measured wind speed (blue circles).
Case study of 27. Ocotber 2013 (UTC-time). Theoretical power curve (black) with measured power output from
WTG08 as a function of forecast wind speed (red circles) . . .
The number of measurements observed for a given wind speed
and direction in the period 1. May 2013 - 31. April 2014
(UTC-time), using hourly averaged data. . . . . . . . . . . .
The standard deviation of the observed power output for a
given wind speed and direction in the period 1. May 2013 31. April 2014 (UTC-time), using hourly averaged data. . . .
The mean (a) and median (b) power output observed for a
given wind speed and direction in the period 1. May 2013 31. April 2014 (UTC-time), using hourly averaged data. Last
row displays the corresponding power curve for the WTG08
turbine. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The distribution of mean wind power observed for wind speeds
in the interval between 15 - 20 m/s for winds from southeast
(a) and west (b) in the period 1. May 2013 - 31. April 2014
(UTC-time), using hourly averaged data. . . . . . . . . . . .
The NRMSE of M1, M2, M4 and M7 shown for each month
with the NRMSE for the whole time period given in parenthesis behind the respective model legend. . . . . . . . . . .
xiv
53
53
54
55
55
57
57
58
58
59
60
62
63
66
4.15 The NRMSE of M1, M2, M13 and M16 shown for each month
with the NRMSE for the whole time period given in parenthesis behind the respective model legend. . . . . . . . . . .
4.16 Case study of 27. October 2013 (UTC-time) showing the
predicted power output from the M2 (black), M4 (red) and
M16 (green) models compared to the measured power output
(blue). The NRMSE corresponding to each model for the
whole time period are given in parenthesis. . . . . . . . . . .
xv
67
68
xvi
List of Tables
1.1
1.2
3.1
3.2
3.3
4.1
Time scale of forecast models and their applications [Wang

et al., 2011]. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Some forecasting models for wind speed and power with their
respective time horizon and approach used [Wang et al., 2011].
3
4
Site coordinates . . . . . . . . . . . . . . . . . . . . . . . . . .
Key technical specifications for WAA151 Anemometer [Vaisala
Oyj, 2002a]. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Key technical specifications for WAV151 Wind Vane [Vaisala
Oyj, 2002b]. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
Overview of the performance of persistent models , Markov

chain models and combined Markov chain and persistent models (MC + M2) for various input parameters. . . . . . . . . .
65
xvii
40
42
xviii
Chapter 1
Introduction
1.1
Short-term wind power prediction
According to the latest IPCC-report the dominant cause of the global warming occurring since the mid-20th century is extremely likely to be due to
anthropogenic influence [IPCC, 2014]. Nevertheless the worlds total energy
consumption continues to increase annually, as does the demand for future
energy production, and fossil fuels are by far the most dominant energy
sources used today [BP, 2013]. In order to face the challenging climate
change and at the same time meet the energy needs of tomorrow, reliable
and sustainable alternatives to fossil fuels most be found and developed.
Harvesting the energy from renewable energy sources is considered to be
one of the leading solutions to this challenge [EREC, 2005]. Wind is one such
source with a great future potential. Looking at for instance the theoretical
potential of wind energy on the earth, it has been determined to be in the
range of 110 1012 50 1012 GJ/yr [Rogner et al.]. Taking into account
the technical limitations of the current technology this number is reduced
to 1500 109 5700 109 GJ/yr and taking into account also the practical
limitations such as conflicting land uses and remoteness the potential is
estimated to be about 250 109 1200 109 GJ/yr [Rogner et al.], or in a
more commonly known energy measure 4.2 109 - 200 109 GWh/yr. This is
approximately 34 000 to 162 000 times the electricity produced in Norway
in 2010 (123 630 GWh/yr) [SSB, 2014].
However, there are some challenges connected with harvesting the wind
energy which must be dealt with. One of the greatest is the natural variability of the wind. The global weather system is both complex and dynamic,
making it hard to accurately predict the future wind parameters such as
speed and direction, and great variations may occur over short time scales
[Marquis et al., 2011].
The electricity generation in a region usually consists of a mix between
three types of power plants: slow-start power plants, such as coal and nu1
CHAPTER 1. INTRODUCTION
clear plants, fast-start power plants, such as natural gas and hydroelectric
plant, and variable renewable energy sources such as wind farms and solar
power. At all times these generation units must be controlled by a balancing authority to ensure both a stable electricity supply corresponding to the
consumer demand and to maintain a mixture of electricity generation which
will give the lowest cost possible for the consumers [Yoder et al., 2013].
Being a variable energy resource, wind energy must be used as it is
produced. A higher penetration of wind energy in the grid would therefore
mean a higher degree of uncertainty. This makes it difficult for the balancing
authority to plan how to meet the energy demand, as slow-start power plants
will not be able to react rapidly to changes in the power load and they will
not run as efficiently at lower capacity, while fast-start power plants can
respond more quickly to the changes in demand, but at the cost of usually
being more expensive [Yoder et al., 2013]. A higher degree of uncertainty
in the grid would require more flexibility in the system, such as for instance
keeping more reserves on-line as a backup, which again would lead to higher
operating costs. As noted by Pourmousavi Kani and Ardehali [2011], very
short-term variations (seconds and minutes) in wind power may also cause
voltage and frequency fluctuations, especially in the event of cut-off speed.
Wind energy also causes some economically challenges for the power
production companies, as they have to make estimates of how much power
they have available for sale. Wrong estimates of the power production would
lead to the companies having to buy expensive last minute power to cover
the guaranteed power delivery for underproduction, or having to sell the
power cheap in the case of overproduction.
Accurate wind power predictions is therefore a key element for an increased penetration of wind energy. This would not only make it easier to
implement wind energy into the grid, it would also reduce the operational
cost and make it easier for the power production companies to make more
accurate predictions of how much energy they can sell on the marked.
1.2
Former research
The stochastic nature of wind makes it difficult to develop models which

accurately predicts future wind speed and direction, thus also the future
power output from a wind turbine. As a result, many different approaches
have been tried to find the ideal power prediction model. This is especially
seen in the work by Giebel et al. [2011] which gives an extensive literature
overview of the state of the art short-term wind power prediction models,
based on review of more than 380 journal and conference papers.
Wang et al. [2011] divides the prediction horizon into the following categories:
1. Immediate-short-term: From seconds up to 8 hours ahead forecasts,
2
1.2. FORMER RESEARCH

Time scale
Immediateshort-term
Range
8-hours ahead
Application
- Real-time grid operations
- Regulation actions
Short-term
Day ahead
- Economic load dispatch planning

- Load reasonable decisions
- Operational security in
electricity marked
Long-term
Multiple days ahead
- Maintenance planning
- Operation management
- Optimal operating cost
Table 1.1: Time scale of forecast models and their applications [Wang et al.,
2011].
2. Short-term: A day ahead forecasts,
3. Long-term: Multiple days ahead forecasts,
with the applications for each time horizon given in Table 1.1.
There are mainly two approaches used to when developing wind power
prediction models: (i) a physical approach and (ii) a statistical approach
[Giebel et al., 2011].
The former is a deterministic method which use physical considerations
and input data to make predictions of the future power output [Carpinone
et al., 2010]. These physical models are based on predictions of the lower
atmosphere or numerical weather predictions (NWP) using input data such
as temperature, pressure, surface roughness and obstacles obtained from
weather forecasts [Wang et al., 2011].
The statistical approach predicts the future power output using historical
obtained data, which could also include NWP results, and is a pure mathematical approach which does not consider any of the physical processes of
wind [Carpinone et al., 2010]. The main idea is use these vast amounts of
historical data to find a relationship to the power output. This typically involve time series analysis [Wang et al., 2011] and artificial neural networks,
which often include the use of recursive techniques . For look-ahead times
above 3-6 hours, models using a time series approach are usually outperformed by models involving a NWP [Giebel et al., 2011].
In order to get the best from each approach, a hybrid approach is typically used in most operational and commercial models today [Giebel et al.,
2011]. Table 1.2 shows an overview of some of the wind forecasting models
described by Wang et al. [2011], with their corresponding time horizon and
approach.
3
Model
WPMS
ANEMOS
ARMINES
(A WPPS)
WPPT
Prediktor
Previento
WEPROG
Time horizon
Immediate-short-term
Immediate-short-term, short-term
Immediate-short-term, short-term
Approach
Statistical
Statistical & Physical
Short-term
Short-term
Long-term
Long-term
Statistical
Physical
Table 1.2: Some forecasting models for wind speed and power with their
respective time horizon and approach used [Wang et al., 2011].
Forecast models used to predict more specific events or scenarios of power

production are also being developed. In a recent study Bossavy et al. [2013]
notes the lack of ability of current forecasting models to properly handle extreme situations related to wind generation, being a result of either extreme
weather phenomena or critical periods for power system operation. A ramp
is one such event, defined in the study as a steep and high increase or drop in
power production from a wind farm within a time period of a few hours. The
study proposes a methodology, which used together with numerical weather
prediction ensembles, provides reliable forecasts with greater accuracy regarding climatology
Due to its simplicity and because many natural processes are considered
as Markov processes [Shamshad et al., 2005], Markov chains have become
a popular tool for developing wind power prediction models based on time
series analysis. A Markov chain represents a system which, based on the
input data, calculates the probability of going from one state to another.
The order of the Markov chain decides how many previous time steps influencing the probability distribution of the current state. Such states could
for instance represent a given power output, wind direction or wind speed.
Another advantage using Markov chain is the possibility of not only making
point predictions, but also probabilistic forecasts, , i. e. give information
about how likely it is for a given prediction to occur.
A study by Shamshad et al. [2005] discuss how first and second order Markov chain models can be used for generating synthetic wind speed
time series, which may further be used as input to a wind energy system.
Carpinone et al. [2010] recommends instead to use wind power measurements directly as input to a Markov chain model in making a immediateterm forecast of the power . Argument being that for wind speed forecasts,
the forecast would have to model the wind farm power curve of interest,
take into account individual turbine curves, site orographic characteristics
and wake effects, before finally converting the wind speed forecast into a
4
1.3. PURPOSE OF THE STUDY

power forecast. A process which could amplify the prediction error and
which would be avoided using power data directly.
Instead of giving a specific wind speed forecast or power forecast Yoder
et al. [2013] suggest using Markov chain models for predicting the 1 hour
ahead categorical change in the wind power. In this case the power change
in the last hour, the current wind power and the 20-minute power trend are
used to provide a probabilistic forecast of three states, -1: a negative trend,
0: no change is expected and 1: a positive trend is predicted.
1.3
Purpose of the study
Fakken wind farm on Vannya in Troms county was completed in the end
of 2012 and consist of 18 Vestas V90 3.0MW horizontal axis wind turbines
(HAWT). Annually the farm is expected to produce 138 GWh, enough power
to supply about 7000 households [Troms Kraft AS, 2012].
The main purpose of this study is to use wind and power data recorded
from a reference turbine and a meteorological station at Fakken wind farm
to develop a power prediction model with a 2-hour prediction horizon (defined in this study as a short-term prediction), using the concept of Markov
chains. The model will be tested using various input parameters, such as
the measured power and wind data and forecast wind speed and direction
obtained using numerical weather predictions (NWP).
It is expected that winds blowing from the west are more turbulent than
other directions due to a local mountain range causing a weather phenomena
known as mountain waves. It has also been suggested that the mountain
range between Norway and Sweden to the southeast, causing regional mountain waves, may also influence the power production on Fakken wind farm.
An analysis of the statistical properties of measured power and wind data
will therefore also be given and related to the topography and surroundings
of the wind farm site.
1.4
Structure of the study
This study is divided into 6 chapters. First chapter provided the motivation
and purpose of this study. Wind theory and statistical theory are given
in chapter 2, while information of the site and data collection are given in
chapter 3. This is also where the prediction models are described. Results
are shown in chapter 4 and discussed in chapter 5. In chapter 6 the study
is concluded together with suggestions for further research.
Chapter 2
Theory
2.1
Source of wind energy
Wind energy, or simply wind, is the kinetic energy of air in motion. As

stated by [Andrews and Jelley, 2007], the main source for wind energy is
solar radiation. Radiation incident on the earth from the sun will mainly
be absorbed by the earths surface, heating it, while the rest of the radiation
will either heat up the atmosphere or be reflected by clouds, aerosols and
gases in the atmosphere [Houghton, 2009]. Due to the shape of the earth,
incident radiation hits the earth at different angles. At equatorial latitudes
the incoming radiation will be close to perpendicular to the surface, while at
higher latitudes, the same amount of radiation will be spread over a much
greater surface area. As a result, the earth will in general receive a lot more
energy at lower latitudes than higher latitudes, which results in a unbalance
in the energy system. In addition to this effect, the heat capacity of the
various materials differ from each other, thus the energy absorbed by the
surface itself will also vary. This leads to temperature variations, which
again leads to pressure gradients and in order to balance the system gas
molecules in the air will start to move, causing convection.
Figure 2.1 shows an illustration of a simple convection cell. As air is
heated up, it becomes less dense and the pressure decreases, creating a low
pressure as can be seen to the right hand side in the figure. The heated air
will start to rise in the atmosphere due to being less dense than nearby air
parcels, but as it rises higher up in the atmosphere it will start to cool down
again. In addition, the gravitational force acting on the air parcels decreases
with height, i. e. air parcels with same temperature at two different heights
will have different pressure forces acting on them. This causes pressure
gradients, and high pressured air molecules next to the low pressure will
start to move towards the low pressure. Simultaneously the air molecules
from the low pressure will now start moving towards the illustrated high
pressure as it is cooled down.
7
CHAPTER 2. THEORY
Figure 2.1: A simple convection cell, with low pressure denoted by L and
high pressure by H.
2.2
Global wind patterns
As mentioned in the previous section the heat energy from the sun is spread
unevenly across the earths surface causing unbalance in the energy system of
the earth. At the same time the earth continuously radiates thermal energy
back into space. This results in a negative energy budget at the poles and
at higher latitudes and a positive budget at lower latitudes [Emeis, 2013].
In order to balance this system, the earth transports vast amounts of heat
from equatorial latitudes to the poles. This complex heat transportation
system makes up the global climate and consists of circulations in the oceans
and atmosphere as well as direct transportation of energy through heat
conduction [Emeis, 2013].
There are three main convection cells making up the global wind system
[Boyle, 2004], seen in Figure 2.2. Description of these cells will only be
given for the northern hemisphere, as they are mirrored for the Southern
Hemisphere.
The first cell, the Hadley cell, goes from about 0 degrees to about 30
degrees latitude. At equator the earths surface consist to a large extent of
oceans. Being also the latitude of which most solar energy is incident on a
given surface area, a low pressure belt is created consisting of a cloudy, warm
and humid climate. This is the result of huge amounts of water evaporating
from the oceans to the atmosphere. As the air is heated up the now humid air
will expand and rise up into the atmosphere until it reaches the tropopause
[Boyle, 2004].
8
2.2. GLOBAL WIND PATTERNS
Figure 2.2: The global wind circulation [Guido, 2008].
While the air rises up, it cools and much of the water content in the air
will start to condensate into clouds. These clouds will work as a blanket
covering the equator and reflect the incoming solar radiation, thus actually
contributing to less solar radiation reaching the surface at equator. As the air
rises to the tropopause, the air will move towards the poles, i.e. northwards
in the Northern Hemisphere and southwards in the Southern Hemisphere
[Boyle, 2004]. The air will keep on cooling until it reaches about 30 degrees
latitudes. At this point the air has become so dense that it will start sinking
down towards the surface. This air is now dry, cold and dense, creating
a high pressure belt, as seen in Figure 2.2, being why many of the earths
deserts are found at 30 degrees latitudes. Parts of the air at the surface will
now be forced towards the low pressure belt at equator, completing the air
circulation between 0 and 30 degrees latitudes. This is known as the Hadley
cell.
The remaining air will be forced towards the poles, and will continue to
move until it reaches the polar front. At this point the lower latitude air will
be warmer than that which comes from the poles, thus lower latitude air rises
up to about 60 degrees latitude, creating another low pressure belt around
the planet. Again one part of this air will move towards equator, more
specifically towards 30 degrees latitude, completing the second convection
cell, i.e. the Ferrel cell. The second part will move towards the poles, until it
9
CHAPTER 2. THEORY
cool down at higher latitudes, which lead to the third and last cell, namely
the Polar cell [Boyle, 2004]. This is the reason for the high pressure at
the poles, and since the air at high latitudes are much colder than that of
lower latitudes, it can hold less water. These polar latitudes are, like the 30
degrees latitude, also characterized by desert areas.
However, as stated by Emeis [2013], these cells would only create a meridional wind pattern. In order to get a more realistic picture of the global wind
patterns, the earths rotation must be included. The earth is rotating from
west to east, or as seen from the North Pole, counter-clockwise. As a result of this, the winds moving towards the poles will get a westerly velocity
component, while the winds moving towards the equator will get a easterly
velocity component. This is the effect of a pseudoforce known as the Coriolis effect. These components give rise to the well known wind patterns as
westerlies, polar easterlies and the trade winds, seen in Figure 2.2.
2.3
Vertical wind speed profile
When discussing the vertical wind speed profile with respect to wind turbines
Emeis [2013] first considers a flat horizontally homogeneous atmospheric
boundary layer (ABL). This layer is the lower part of the Troposphere and
the lowest layer of the atmosphere. According to Emeis [2013] the atmospheric boundary layer can be divided into three principal types
1. Neutral boundary layer: in this layer the heat flux at the lower surface
is negligible and the layer are instead dominated by dynamical shear
forces.
2. Unstable boundary layer: this layer is dominated by heat input from
below causing a convective (unstable) boundary layer.
3. Stable boundary layer: when the atmosphere is cooled from below the
boundary layer is said to be stable.
Figure 2.3 shows the vertical stratification of a ABL with the Ekman,
Prandtl and roughness sublayers illustrated with their corresponding order
of heights. Two wind turbines are also illustrated as a reference to the layer
heights. Both turbines have a hub height of 80 meter and a rotor diameter
of 90 meter, which is the dimensions given for the V90-3.0MW located at
Fakken wind farm [Vestas Wind Systems A/S, 2013]. The left hand side is
illustrated relative to a zi = 1000 m, which will result in the Prandtl layer
going up to 100 meters, and the right hand side wind turbine relative to zi
= 2500 m. It should be noted that in the left hand case the tip height of
the wind turbine extends into the Ekman layer, whereas for the right hand
case the tip height is within the Prandtl layer with good margins.
10
2.3. VERTICAL WIND SPEED PROFILE
Free troposphere
zi= o(1000) - o(2500) m
Ekman layer
o(0.1zi)
Prandtl layer
Roughness layer
o(10z0)
Z0
Figure 2.3: Vertical layers of the atmospheric buondary layer over flat homogenous terrain.
As meteorology is not the main focus of this study, only a short summary
of the various layers within the ABL and the corresponding vertical wind
profile laws will be given in the following sections. More detailed information
on this subject is given in [Emeis, 2013] and the pre-study for this thesis
[Jacobsen, 2013].
The Prandtl layer, also known as the surface layer, is the layer between
the Ekman and roughness layers. As seen in Figure 2.3 this layer usually
makes up 10 % of the ABL. Since the Prandtl is so close to the surface,
the affect from the Coriolis force can be neglected and meteorologically the
layer is defined as where the turbulent vertical fluxes of moisture, heat and
momentum differs from their surface values with only 10 % [Emeis, 2013].
In the Ekman layer the wind profile is influenced by three forces, (i) the
Coriolis force, (ii) the pressure gradient force and (iii) the frictional forces.
2.3.1
Logarithmic law
The vertical wind speed gradient, also known as the wind shear, in the
Prandtl layer in the case of neutral stratification is given as Emeis [2013]
u
u
u
=
=
,
z
l
z
(2.1)
where u is the wind speed, z the height, u the friction velocity, l the mixing
length and = 0.4 is the K
arm
an constant
11
CHAPTER 2. THEORY
It is then possible to derive an expression for the wind speed as a function
of height by integrating Equation 2.1 from the lower height z (roughness
length), where the wind speed is assumed to vanish near the ground, up to
a height z within the Prandtl layer
Z
u
dz
z0 z
du =
0
u
(ln z ln z0 )
u
z
ln
z0
u(z) =
1
dz
z0 z
(2.2)
z
u
ln .
z0
However, rarely is the thermal stratification of the Prandtl layer found

to be absolutely neutral. For unstable and stable stratifications a correction
function must therefore be added to Equation 2.2 so it becomes [Emeis,
2013]
u
u(z) =

z
z
ln
m
,
z0
L
(2.3)
where L is the Obukhov length defined as

L =
v u3
.
kg v0 w0
(2.4)
Here u is the friction velocity, v the potential temperature, g the gravitational acceleration and v w is the virtual heat flux. The Obukhov length
is negative for unstable stratification and positive for stable stratification.
More details regarding the Obukhov length can be found in [Emeis, 2013].
In the case of unstable stratification the correction function becomes

1 + x2
1+x
+ ln
2 arctan(x) + ,
(2.5)
m = 2 ln
2
2
2
where x = ( bz/L )/ [Paulson, 1970] and b = [Emeis, 2013]. For
12
2.4. DISTRIBUTION OF WIND

stable stratification the correction function becomes
for 0 < z/L 0.5
a(z/L )
m = A(z/L ) + BC/D + ...
B (z/L C/D) exp (D(z/L ))

for 0.5 z/L 7,
(2.6)
with a = 5, A = 1, B = 2/3, C = 5 and D = 0.35 [Emeis, 2013].
2.3.2
Power law
The power law is an empirical law used to describe the vertical wind profile. It is often preferred to the logarithmic law due to its mathematically
simplicity and it is defined as [Emeis, 2013]
a
z
u(z) = u(zr )
.
(2.7)
zr
The Hellman exponent a depends on the surface roughness and the thermal
stability of the Prandtl layer and zr is the reference height.
2.4
2.4.1
Distribution of wind
Weibull distribution
The Weibull distribution is a common function used to model how the occurrences of wind speeds are distributed [Seguro and Lambert, 2000]. In a early
study by Justus et al. [1976] this distribution was seen to fit well with the
observed wind speed data measured at various sites across the continental
United States.
For a random Weibull distributed variable U W (a, b) , Conradsen
et al. [1984] describes the following statistical characteristics:
Probability density function

b u b1
u b
f (u; a, b) =
exp
,
a a
a
Cumulative distribution function

u b
,
F (u; a, b) = 1 exp
a
u 0,
u 0,
(2.8)
(2.9)
Mean

1
E(U ) = = a 1 +
,
b
13
(2.10)
CHAPTER 2. THEORY
0.25
a = 6, b = 1
a = 6, b = 2
a = 6, b = 3.6
0.2
f(u)
0.15
0.1
0.05
0
0
10
15
m/s
20
25
30
Figure 2.4: Weibull probability density function as a function of wind speed

for various values of b.
Variance

2
1
V ar(U ) = 2 = a2 1 +
2 1 +
,
b
b
(2.11)
where () is the gamma function. For b = 1 the distribution is an exponential distribution with mean value a, while b = 2 gives a Rayleigh distribution. When b = 3.6 the distribution approaches the Gaussian distribution
[Conradsen et al., 1984]. Figure 2.4 show some examples of the probability
density function with various parameter values.
Seguro and Lambert [2000] describes three methods for estimating a
and b of the Weibull distribution: (i) the maximum likelihood method, (ii)
the modified maximum likelihood method and (iii) the graphical method.
Methods (i) and (ii) are described as follows:
The maximum likelihood method

Consider a set of n measurements of non-zero wind speeds u, where ui is the
wind speed at time step i. The parameters a and b can then be estimated
14
2.4. DISTRIBUTION OF WIND

as
n
a=
1X b
ui
n
!1
b
(2.12)
i=1
Pn
Pn
1
b
i=1 ln(ui )
i=1 ui ln(ui )
Pn
b=
,
b
n
i=1 ui
(2.13)
where Equation 2.13 must be solved before Equation 2.12, using an iterative
procedure of which a suitable initial guess is b = 2 [Seguro and Lambert,
2000]. This is known as the maximum likelihood method.
The modified maximum likelihood method
In the case that the wind speed data are given in frequency distribution
format, i. e. the data are sorted into equally sized bins corresponding to
wind speed intervals, such as for instance used when representing data in
a histogram, a modified version of the maximum likelihood method can be
applied instead using the following equations [Seguro and Lambert, 2000]
n
a=
X
1
ubi P (ui )
P (u 0)
!1
b
(2.14)
i=1
Pn
Pn
1
b
i=1 ui ln(ui )P (ui )
i=1 ln(ui )P (ui )
Pn
b=
,
b
P (u 0)
i=1 ui P (ui )
(2.15)
where n now is the number of bins which the data are sorted into and u is
still the wind speed, while ui is the wind speed central to bin i. P (ui ) is the
probability of wind speed u falling within the intervals of bin i and P (u 0)
is the probability that the wind speed is equal or greater than zero. Like
before, Equation 2.15 must be solved iteratively before an explicitly solution
for a using Equation 2.14 can be obtained.
The Weibull parameters a and b are related to the average wind speed
by [Seguro and Lambert, 2000]

1
u
= a 1 +
,
(2.16)
b
where () is the gamma function.
2.4.2
Alternative distributions
In a recent study by Drobinski and Coulais [2012] the authors questions the
general suitability of the Weibull distribution to describe the wind speed
distribution and thus the wind power evaluation at a site. In their article
data from four different sites in France were evaluated. These showed that
15
CHAPTER 2. THEORY
for the site with relatively flat terrain and isotropic winds, the distribution
of the wind could be fairly well described using the Weibull distribution.
For the three remaining sites with more complex terrain and anisotropic
wind however, the Weibull distribution did not very well describe the wind
distribution. According to Drobinski and Coulais [2012] the reason for this
was assumed to be that the area for the three remaining sites had two wind
regimes: (1) random flow: wind components have zero means and same
variance and (2) channeled flow: wind components have different means.
The wind statistics of the first regime is well described by the Rayleigh
distribution and the second regime by the Rice distribution. A RayleighRice distribution, given by Equation 2.17 which takes into account the wind
components instead of only the wind speed was therefore proposed to be
used. This gave a much lower relative error for the complex sites compared
to the Weibull distribution. This may suggest that how well the Weibull
distribution fits the wind speed distribution varies with features such as the
complexity of the terrain and whether the wind is isotropic or anisotropic.

M
M2
pM (M ; , , ) = exp 2
Z
2

2
M
+ (1 ) exp 2 I0
,
2
2
2
(2.17)
where M = u2 + v 2 is the wind velocity with wind speed components u

and v, I0 () is the modified Bessel function of the first kind and zero order,
R the weight 2corresponding to channeled flow events occurrence and Z =

pM (M ; , , )dM is the normalization factor [Drobinski and Coulais,
2012].
2.5
2.5.1
Wind power production

Energy in the wind
Wind energy is the kinetic energy of air in motion. As described by Boyle

(2004) this kinetic energy is defined as
1
Ek = mu2 ,
2
(2.18)
where m is the mass of moving air and u its velocity, in kilograms and meters
per second respectively.
Considering a cross-sectional area A being perpendicular to the wind
direction of u, the volumetric flow rate of the air Q, i.e. the volume of air V
flowing through A per unit time t, can be expressed as
16
2.5. WIND POWER PRODUCTION
Q=
dV
= uA.
dt
(2.19)
Starting with the definition of power it is possible to derive an expression

for the power available in the wind.
P =
=
=
=
=
=
=
=
dW
dt
Ek
t
1
2
2 mu
t
1 m 2V
u
2 t V
1 m 2V
u
2V
t
1 2
u Q
2
1 2
u (uA)
2
1
Au3 ,
2
(2.20)
where is the density of the air. In other words, the power in the wind is
proportional to the third order of the wind speed u.
However, as noted by Andrews and Jelley [2007], all of this power cannot
be extracted using a wind turbine. This has to do with the fact that some
kinetic energy is needed in order to maintain an air flow through the wind
turbine. The wind turbine will convert the kinetic energy to mechanical
energy, thus slowing down the wind speed. If the wind is slowed down
completely, the air flow through the turbine would cease and so would the
power extraction.
Betz limit
Using momentum theory it is possible to make an estimate of the maximum
efficiency of power extraction from the wind using a wind turbine [Andrews
and Jelley, 2007]. First consider the stream-tube seen in Figure 2.5. Starting
with speed u the wind flow upstream pass through a cross-sectional area A
As it reaches the wind turbine, the area of the steam-tube of air increases
to A and as a result the wind speed will be decreased to u . Some of the
kinetic energy of the air will be converted to mechanical energy in the rotor,
reducing the wind speed further in addition to the effect from the increased
area of the stream-tube downstream.
17
CHAPTER 2. THEORY
There are two ways to express the power extracted by the wind turbine
[Andrews and Jelley, 2007]. First as the loss of kinetic energy per unit time
as the wind flow through the turbine
P = Pupstream Pdownstream
Ek0 Ek2
=
dt
dt
1 dm 2 1 dm 2
u
u
=
2 dt 0 2 dt 2

1 dm 2
=
u0 u22 .
2 dt
(2.21)
Second expression is obtained by looking at the change of momentum.

Consider the rotor as a actuator disc, i.e. a thin disc that extracts energy.
As the wind speed drop when the wind flow passes through the turbine, a
pressure drop will occur [Andrews and Jelley, 2007]. Bernoullis theorem
states that for an ideal fluid, the energy must be conserved. Assuming the
air is an ideal fluid, it is possible to use Bernoullis equation to find the
trust, T, exerted on the turbine blades which again will lead to the power
extracted. The Trust is given by
dm
(u)
dt
dm
=
(u0 u2 ),
dt
T =u
(2.22)
where T is the thrust force and dm/ dt the rate of which the mass of air flow
through the stream-tube. The extracted power as a function of the thrust
can now be expressed as
P = T u1
dm
(u0 u2 ) u1 .
=
dt
(2.23)
Combining Equation (2.21) and Equation (2.23), an expression for u

and u can be obtained by

dm
1 dm 2
(u0 u2 ) u1 =
u0 u22
dt
2 dt

1 2
(u0 u2 ) u1 =
u0 u22
2
1
u1 = (u0 + u2 )
2
u2 = 2u1 u0 .
18
(2.24)

and the mass flow per second can be rewritten as [Andrews and Jelley, 2007]
dm
= uA = u1 A1 .
dt
(2.25)
Using Equations (2.24) and (2.25), Equation (2.23) becomes

P = (u1 A1 ) (u0 (2u1 u0 )) u1
= 2u21 A1 (u0 u1 ).
Define an induction factor a, so that
u0 u1
a=
,
u0
and a power coefficient Cp as
Cp =
(2.26)
(2.27)
P
1
3
2 u0 A1
(2.28)
2
= 4a(1 a) ,
which is the ratio between the power in the wind and the power extracted
by the turbine. It is then possible to find the solution for a which will yield
the maximum power extraction as follows
dCp
=0
dt

d
4(1 a)2 = 0
dt
12a2 16a + 4 = 0,
(2.29)
which is on the form of a quadric equation that can be solved using the
quadric formula resulting in the two solutions
(
1
(2.30)
a= 1 .
3
For a = 1, P = 0, and the maximum power is therefore found for the

solution a = 1/3, or the equivalent u1 = (2/3)u0 . Using Equation (2.28)
the limit for the power coefficient can be found as

!
1 3
1
1 2
Pmax = u0 A1 4
1
2
3
3

(2.31)
1
16
= u30 A1
2
27
= Pwind Cp .
In other words, the maximum ratio of power which can be extracted by the
wind turbine, Pmax , of the power in the wind incident on the turbine, Pwind ,
is given by the power coefficient Cp . This limit is called Betz limit and is
approximately 59 % [Andrews and Jelley, 2007].
19
CHAPTER 2. THEORY
A2
A1
A0
u0
u2
u1
Upstream
Wind turbine
Downstream
Figure 2.5: Airflow through a HAWT.
2.5.2
Wind turbine blade design
The blades of a modern wind turbine are most commonly shaped like an
airfoil, reason being that the shape allows for efficient extraction of the
energy in the wind [Mathew, 2006]. It is therefore natural to first look
at how lift is created on an airfoil and a good review on this is given by
Babinsky [2003].
Consider the airfoil seen in Figure 2.6. The white lines seen are smoke
particles flowing along a lifting airfoil section [Babinsky, 2003]. To the left
in Figure 2.6a the streamlines are seen to be undisturbed by the airfoil,
however as the smoke particles approaches the airfoil the streamlines curves
around the airfoil as seen in Figure 2.6b and Figure 2.6c. This happens
because the shape of the airfoil induces curvature into the flowfield and the
flow will follow the local curvature of the upper and lower surfaces, which
again is a result from friction between the airfoil and the particles.
As noted by [Babinsky, 2003], in order for the streamlines to be curved,
there must be a force acting on the particles normal to the direction of
motion. This force origins from pressure gradients and for the case seen
in Figure 2.6 the pressure above the airfoil will be lower than the normal
atmospheric pressure, while at the lower side of the airfoil the pressure will
be greater. These pressure gradients are greater closer to the surface, as
seen in the figure. Note also that the air particles moving above the airfoil
moves faster than the ones below. This can be explained by Bernoullis law
which states that the decrease in pressure of an inviscid flow will result in
20
(a)
(b)
(c)
Figure 2.6: Air particles flowing along an airfoil [Babinsky, 2003].
(a) Low angle of attack
(b) High angle of attack
(c) Stalled flow
Figure 2.7: Streamlines flowing along an airfoil with increased angle of attack
[Babinsky, 2003].
an increased speed of the fluid, vice versa [Giancoli, 2005].

Figure 2.7 shows how the streamlines around an airfoil are affected by
the angle of attack. For low angle of attack, as seen in Figure 2.7a, the
streamlines are less densely packed around the surface of the airfoil compared
to Figure 2.7b. This implies that there is more lift generated for a higher
angle of attack. If the angle is increased too much however, the flow above
the airfoil will no longer be able to follow the sharp curvature of the airfoil,
resulting in the sharp drop-off in the lift force seen in Figure 2.7c [Babinsky,
2003].
Mathew [2006] divides the force exerted on the wind turbine blade into
two components: The lift force L, which is the force perpendicular to the
direction of the undisturbed flow, and the drag force D, which is the force
exerted in the same direction as the undisturbed flow. These forces are seen
in Figure 2.8b, which illustrates a section of one of the blades at a length
r from the rotor center of the wind turbine seen in Figure 2.8a. The wind
incident on the turbine has a speed u and the blade a rotational speed v,
perpendicular to the direction of the air flow.
The lift and drag forces are given as [Mathew, 2006]
1
L = CL a A1 u21 ,
2
21
(2.32)
CHAPTER 2. THEORY
and
1
D = CD a A1 u21 ,
(2.33)
2
respectively, where CL and CD are the corresponding lift and drag coefficients, a is the air density, A1 the cross-sectional area of the turbine and
u the incident wind speed.
For a turbine blade with maximum radius R, the speed of the blade at a
radius r from the center of the rotor is given by Andrews and Jelley [2007]
to be
rvtip
v=
.
(2.34)
R
Using trigonometric properties, the following relation can be derived from
Figure 2.8b
u1
tan() =
,
(2.35)
v
where is the angle between v and the resultant speed of the air relative to
the blade, u .
Neglecting the drag force, Andrews and Jelley [2007] derives an expression for the power as a function of the lift force as follows
P = Fv
= L(sin())v
u1
tan()
= L(cos()u1
= L(sin())
(2.36)
= T u1 ,
which implies that the power developed equals the power delivered by the
thrust force, as was seen in Equation 2.23.
Another critical parameter which should also be considered is the tipspeed ratio , defined by Andrews and Jelley [2007] as
=
vtip
,
u0
(2.37)
where vtip is the speed of the blade tip and u0 the wind speed incident on
the cross-sectional area A0 , as seen in Figure 2.5.
Combining Equations 2.37 and 2.34 with Equation 2.35 yields
tan() =
u1
2R
=
,
v
3r
(2.38)
where the Betz condition u1 = (2/3)u0 has been used. From this result
it is possible to observe that the angle only depends on the tip-speed
ratio for a given radius r. As increased radius r implies that the angle
will decrease, modern wind turbine blades are design with a twist which
22

increases with decreasing r. By doing so the optimal angle of attack are
obtained over the whole turbine blade. Modern blades are also design so
that the width W decreases with increased radius r. This is to make the
lift components generate the required thrust for the Betz condition to be
maintained [Andrews and Jelley, 2007].
2.5.3
Horizontal axis wind turbine design
Like the majority of wind turbine designs today, the wind turbines installed
at Fakken wind farm are HAWTs [Andrews and Jelley, 2007]. A detailed
illustration of the various components included in a modern HAWT is shown
in Figure 2.9. The rotor consists of a hub, typically with three turbine blades
shaped like airfoils attached to it. This is also the case for the Vestas V90
3.0 MW wind turbines installed at Fakken wind farm, which have rotor
diameters of 90 meter [Vestas Wind Systems A/S, 2013]. Each blade can
be rotated, known as pitching the blade, in order to most efficiently harvest
the wind energy and when the wind speed is too high, reduce the fatigue on
the blade.
On top of the tower is the nacelle which encloses most of the important
components of a wind turbine. As can be seen in the illustration, the rotor
is attached to the nacelle. Closest to the rotor lies the break whose main
function is to be a backup to the breaking effect obtained by pitching the
blade and work as a parking break when conducting maintenance. A lowspeed shaft is used to transfer the mechanical energy from the rotor, by a
gear box, to a high-speed shaft attached to a generator. For large turbines,
the low-speed shaft will typically have 20 revolutions per minute (rpm) and
by using the gear box this speed can be increased to about 1200 - 1800
rpm for the high-speed shaft [Schubel and Crossley, 2012]. The increased
speed makes it possible to convert the mechanical energy from the rotor into
electricity using the generator.
Next to the generator is a heat exchanger, which keeps the generator
cool, and a control unit. As the weather conditions such as wind speed and
direction are constantly changing it is very important to find the most ideal
settings for the wind turbine to ensure the highest electricity production.
The controller have a computer system which takes care of this. This system
can run self-diagnostic tests, start and stop the turbine as well as make
adjustment with regards to the pitching of the blades and the direction of
the nacelle using the yaw drive.
For the controller to find the most ideal settings, it requires some input
data. An anemometer and a wind vane are therefore placed on top of the
nacelle, passing along data regarding the wind speed and wind direction
respectively.
The two last components seen in the wind turbine illustration in Figure
2.9 are the yaw drive and the tower. The nacelle can be turned using the yaw
23
CHAPTER 2. THEORY
tat
ion
Ro
u1
r
(a) HAWT.
L
v
u1
(b) Turbine blade section.
Figure 2.8: Wind incident on a horizontal axis wind turbine (a), with a
section of a turbine blade at length r from the rotor center given in (b).
24
1. Rotor
2. Pitch drive
3. Nacelle
4. Brake
5. Low-speed shaft
6. Gear box
7. High-speed shaft
8. Generator
9. Heat exchanger
10. Controller
11. Anemometer
12. Wind vane
13. Yaw drive
14. Tower
Figure 2.9: Design of a modern HAWT [Schubel and Crossley, 2012].
25
CHAPTER 2. THEORY
3500
Rated wind speed
3000
Cutout wind speed
Rated power
Power [kW]
2500
2000
1500
1000
Cutin wind speed
500
0
0
10
15
Wind speed [m/s]
20
25
30
Figure 2.10: Power curve showing the relation between output power and
wind speed for a Vestas V90 3.0MW HAWT. Values in graph are taken from
[Vestas Wind Systems A/S, 2013].
drive, keeping the rotor facing the wind. As wind speed generally increases
with height, the height of the tower corresponds to how much of the wind
energy the turbine are able to capture. The wind turbines at Fakken have
a hub height of 80 meters [Vestas Wind Systems A/S, 2013].
2.5.4
Power curve
The power curve of a wind turbine shows the expected output power of the
turbine as a function of wind speed. This curve is unique for every type
of wind turbine and can be used to find the most suitable type of turbine
for a given site. A typical power curve is seen in Figure 2.10, which shows
the power curve for the Vestas V90-3. 0 MW HAWT [Vestas Wind Systems
A/S, 2013].
The rotor starts rotating when the wind speed reaches the cut-in wind
speed at 3.5 m/s. At 15 m/s, rated wind speed, the maximum output power
is generated by the turbine and this continues up to 25 m/s which is the
cut-out wind speed. This is a safety limit set to avoid great loads and fatigue
on the turbine. At this point the rotor is slowed down either by changing
the pitch of the blades which will reduce the angle of attack, by the design
of the blade or by applying the shaft brake [Andrews and Jelley, 2007].
26
2.6. TOPOGRAPHIC FEATURES
2.6
2.6.1
Topographic features
Wake effects
When wind traverse through a wind turbine some of its kinetic energy will
be converted to electrical energy. As a result, the wind at the lee side of the
turbine will have a lower speed and energy content than the wind upstream
of the turbine [Koch et al., 2005]. In addition to a reduced speed, the wind
downstream of the turbine will also be turbulent. This downstream wind is
known as the wake of the turbine and according to Gonzalez-Longatt et al.
[2012] these two features are the main effects of a wake.
For a wind farm, the reduced wind speed downstream will affect the
energy production from downwind turbines negatively. These turbines are
said to be shadowed by the wake producing turbine. According to Koch
et al. [2005] the total energy production from a wind farm is reduced by a
few percentage as a result of this wake effect. Additionally the increased
turbulence of the wind may result in increased dynamic mechanical loads
on the downward turbines [Gonz
alez-Longatt et al., 2012].
In order to maximize the energy production from a wind farm, thus
increasing the efficiency of the farm, it is very important to understand and
minimize the wake effects when designing wind farms. A numerous selection
of both simple and complex models have been developed to simulate these
wake effects, of which two of them are described by Koch et al. [2005] and
Gonzalez-Longatt et al. [2012].
2.6.2
Mountain waves
Mountain waves are a feature of mesoscale meteorology which occur when

stable stratified air is forced to flow over sinusoidally varying surface topography, for instance mountains [Holton and Hakim, 2012]. These mountain
waves are divided into two categories: (i) the vertically propagating and (ii)
the vertically decaying.
Figure 2.11 shows vertically propagating mountain waves. A upstream
phase tilt occur when the streamlines flow over the mountain ridge which
propagates vertically. In this case the mean cross mountain wind speed is not
increasing significantly and the stability of the stratification increases with
height. The pattern is periodic with height and if enough moisture is present
orographic clouds may develop in the regions where the streamlines are
displaced, both upstream and downstream of the ridge [Holton and Hakim,
2012] as can be observed in Figure 2.11.
However, if the stability of the layers decreases strongly with height and
the mean cross mountain wind speed increases strongly with height, the
vertically propagating waves in the lower layer will be reflected when they
reach the upper layer. If this occur repeatedly the result becomes so called
trapped lee waves as can be seen in Figure 2.12.
27
CHAPTER 2. THEORY
Under some conditions partial reflection of the vertical propagating waves
can also produce strong surface winds along the lee slopes of mountain ranges
[Holton and Hakim, 2012].
2.7
Numerical weather prediction model
All forecast data used in this study origins from the numerical weather prediction (NWP) model AROME-Norway. The AROME-model is relatively
new, becoming operational at Meteo-France in the end of 2008 [Seity et al.,
2011] and the Norwegian Meteorology Institute started using the Norwegian
version AROME-Norway for their official weather forecasts at 1. October,
2013 [Norwegian Meteorological Institute, 2013].
AROME-Norway has a resolution of 2.5km with a grid size of 750x960
and 65 vertical layers [Aspelien and Kltzow, 2013], compared to the previously used HIRLAM8 and HIRLAM12 models with 8 and 12 km resolution
respectively [Norwegian Meteorological Institute, 2008]. One benefit of the
new model is naturally the better resolution, which among others will give
large improvement for precipitation forecasts.
Forecasts are issued four times a day at: 00UTC, 06UTC, 12UTC and
18UTC, each with a 66 hour forecast horizon. These forecast are provided
as point forecasts for every integer hour.
2.8
2.8.1
General statistics
Mean
For a random variable X, taking on values x, and with a probability distribution given by f (x), Walpole et al. [2007] defines the expected value or
mean of X as
X
= E(X) =
xf (x),
(2.39)
x
for discrete X, and

Z
= E(X) =
xf (x)dx,
(2.40)
for continuous X.
Assuming a random sample {X1 , X2 , ..., Xn } where the data in the sample are realizations of n independent and identically distributed observations
of the generic random variable X, the mean of X can be estimated as
n
X (n) ,
1X
Xi ,
n
(2.41)
i=1
known as the sample mean, or mean-estimator function [Stark and Woods,

2012].
28
2.8. GENERAL STATISTICS
Figure 2.11: Vertically propagating mountain waves [Muller, 2014].
Figure 2.12: Vertically decaying mountain waves [Muller, 2014].
29
CHAPTER 2. THEORY
2.8.2
Variance
Walpole et al. [2007] defines the variance of a random variable X with probability distribution f (x) and mean as
h
i X
2 = E (X )2 =
(x )2 f (x),
(2.42)
for discrete X, and

h
i Z
2 = E (X )2 =
(x )2 f (x)dx,
(2.43)
for continuous X, where is known as the standard deviation of X, usually

abbreviated STD.
The variance-estimator function or sample variance is equivalently defined by Stark and Woods [2012] as
n
X
(n) ,
1 X
(Xi
X (n))2 ,
n1
(2.44)
i=1
2.8.3
Cumulative distribution function
For a random variable X, Leon-Garcia [1994] defines the cumulative distribution function (cdf) as
FX (x) = P (X x)
for
< x < ,
(2.45)
which means the probability of a random variable X taking on a value in

the set (, x]. The cdf has the following properties:
1. 0 FX (x) 1,
2. lim FX (x) = 1,
x
3.
lim FX (x) = 0,
4. FX (x) is a nondecreasing function of x, which means if a < b then,

FX (a) FX (b),
5. FX (x) is continuous from the right, which means for h > 0, FX (b) =
lim FX (b + h) = FX (b+ ).
x
30
2.8.4
Probability density function
Leon-Garcia [1994] defines the probability density function of a random

variable X, if it exists, as
fX (x) =
dFX (x)
,
dx
(2.46)
with the following properties:

1. fX (x) 0.
Z
2. P (a X b) =
fX (x)dx.
a
3. FX (x) =
fX (t)dt.
fX (t)dt = 1.
4.
2.8.5
Median
For a distribution of a random variable X, the median of the distribution is

given by Hogg and Craig [1978] as the value of x fulfilling the following two
conditions
1
P (X < x) ,
(2.47)
2
1
P (X x) .
(2.48)
2
2.8.6
Error measures
A standardization of the performance evaluation of short-term wind power

prediction models are given by Madsen et al. [2005]. Here the prediction
error is defined as the difference between measured and predicted value
e(t + k|t) = P (t + k) P (t + k|t),
(2.49)
where t + k is the lead time (usually in hours) and P and P are the true
and predicted power respectively. Dividing the prediction error e on the
installed capacity Pinst will give the normalized prediction error,

1
(t + k|t) =
P (t + k) P (t + k|t) .
(2.50)
Pinst
The prediction error can be decomposed into a systematic error e and
a random error e so that
e = e + e ,
(2.51)
where e is a constant and e is a zero mean random variable.
31
CHAPTER 2. THEORY
Normalized root mean square error
The Root Mean Square Error (RMSE) and the Normalized Root Mean
Square Error(NRMSE) can be defined as
N
1 X 2
e (t + k|t)
N
RM SE(k) =
! 12
,
(2.52)
t=1
and
! 21
N
1 X 2
(t + k|t)
,
N
N RM SE(k) =
(2.53)
t=1
where N is the number of predictions made in the time period evaluated.

These error measures includes contributions from both the systematic and
random error [Madsen et al., 2005].
Model comparison
In order to evaluate and quantify the benefit of using a new model compared
to a reference model, a improvement parameter I can be used, which for a
given lead time is defined as [Madsen et al., 2005]
Iref,EC = 100
ECref (k) EC(k)

(%),
ECref (k)
(2.54)
where EC stands for Evaluation Criterion, for instance the RMSE or NRMSE,
of the new model and the reference model.
2.8.7
Directional statistics
When dealing with directional wind data measurements problems will be

encountered using common statistical features such as mean and variance.
Reason for this being that these features are based upon linear statistics,
thus a new set of statistical features for the non-linear case must be obtained.
For instance, taking the mean of two angles such as 355 and 5 using linear
statistics would result in 180 , i.e. south. However, a more suitable mean
when dealing with wind directions would be 0 , i.e. north which is the
complete opposite.
Directional mean
Jammalamadaka and Sengupta [2001] describes the following method for
finding the directional mean of a dataset D = { , , ..., n }, consisting of
angles measured in radians.
First convert the data into two-dimensional unit vectors r1 , ...rn which
can be represented on a unit circle so that kri k = 1 for i = 1, ..., n, as
32
r sin( )
r cos( )
Figure 2.13: Relation between kartesian and polar coordinates [Jammalamadaka and Sengupta, 2001].
seen in Figure 2.13 and with the relation between the kartesian and polar
coordinates given by
r = (x, y) = (r cos , r sin ) .
(2.55)
A resultant vector R can then be found by summing the n unit vectors

!
n
n
X
X
R=
cos i ,
sin i = (C, S),
(2.56)
i=1
i=1
where kRk = C 2 + S 2 is the length R.

The directional mean can then be found with solutions specified for each
quadrant of the unit-circle as [Jammalamadaka and Sengupta, 2001]

S
arctan C
,
for C > 0, S 0,
for C = 0, S > 0,

2,

S
S
= arctan
= arctan C + , for C < 0,
(2.57)
C
arctan C + 2, for C 0, S < 0,
undefined,
for C = 0, S = 0.
Directional variance
It is often usefull to find a measure for how widely spread the dataset are.
For this purpose Mahan [1991] uses the length R of the resultant vector R.
A large R indicates concentrated data, while a small R indicates that the
data are widely spread, as the data points in many different directions. In
33
CHAPTER 2. THEORY
NORTH
15%
12%
9%
6%
3%
WEST
EAST
Wind speed [m/s]
SOUTH
25 99
20 25
15 20
12.5 15
10 12.5
7.5 10
5 7.5
3.5 5
0 3.5
Figure 2.14: Example of a wind rose
the case of equally dispersed data, R = 0, and no mean can be obtained.

As a result 1 R becomes a measure of the dispersion of the data and the
circular variance becomes
2 = 2(1 R),
(2.58)
where is known as the circular standard deviation, an equivalent to the

linear standard deviation described in Section 2.8.2 [Mahan, 1991].
Circular histogram
A circular histogram, or rose diagram, is basically a histogram where the
bars are replaced by sectors around a circle. Each bar is divided into bins
representing data intervals, with the radius of each bar representing the
percentage of the total data observed within a given sector. The angular
width of the bars are constant [Mardia and Jupp, 2000].
Rose diagram representation of wind direction and speed are known as
wind roses. An example of a wind rose can be seen in Figure 2.14. The circle
is divided into 24 directional sectors, each representing 15 and each sector
are further divided into bins representing different wind speed intervals. The
radius of each sector now represents the percentage number of wind speed
measurements observed for the given sector.
34
2.9. MARKOV CHAINS
2.9
Markov chains
The theory of Markov chains was according to Balzter [2000] developed by

the Russian mathematician Andrei Andreyevich Markov (1856-1922) in his
paper Extension of the Limit Theorems of Probability Theory of a Sum of
Variables Connected in a Chain [Markov, 1907].
Balzter [2000] defines a Markov chain as a stochastic process fulfilling
the Markov property with a discrete state space and a discrete or continuous
parameter space. A stochastic process is a collection of random variables
such that for each index t T ,the index set, X(t) is a random variable [Ross,
2010]. Usually, t is interpreted as time, making X(t) the state of the process
at time t.
For a first-order Markov chain, the Markov property states that the next
state only depends on the present state, which for a discrete case is given as
[Ross, 2010]
Pij = P (Xt+1 = j|Xt = i, ..., X0 = i0 ) = P (Xt+1 = j|Xt = i),
(2.59)
for all t 0 and states i0 , ..., it1 , i, j, where X0 denotes the initial state of
the process and Xt = i that the process is in state i at time t. Pij is the
probability of going from state i to state j, or the transition probability, and
is independent of t. For a second order Markov chain the process depends on
the current state and the immediately preceding state, following the same
pattern for higher orders [Shamshad et al., 2005].
Combining all the transition probabilities between k states, a transition
matrix P of a first-order Markov chain can be defined as [Balzter, 2000]
p1,1 p1,2 p1,k

p2,1 p2,2 p2,k
P = .
..
.. ,
.
.
.
.
.
.
.
pk,1 pk,2 pk,k
(2.60)
where p1,2 is the probability of going from state 1 to state 2. The following
conditions must be fulfilled:
1. pij 0, i, j 1,
2.
Pk
j=1 pij
= 1.
The first condition simply says that all the probabilities must be positive,
while the second condition states that the probability of going from one
state to any other possible state must be equal 1. If it does not equal 1, it
means that all the states are not identified in the transition matrix.
35
CHAPTER 2. THEORY
For a second-order Markov chain, the transition matrix becomes [Shamshad
et al., 2005]
p1,1,1 p1,1,2 p1,1,k

p1,2,1 p1,2,2 p1,2,k
..
..
..
.
.
.
.
.
.
p1,k,1 p1,k,2 p1,k,k
,
P =
(2.61)
p
p
p
2,1,1
2,1,2
2,1,k
p2,2,1 p2,2,2 p2,2,k
.
..
..
..
..
.
.
.
pk,k,1 pk,k,2 pk,k,k
where pijl is the transition probability of going to state l, given that the
current state is j and the previous state was i. The same conditions must
be fulfilled:
1. pijl 0, i, j, l 1,
Pk
2.
l=1 pijl = 1.
It is possible to estimate the transition probabilities of a first- and secondorder Markov chain, using Maximum Likelihood, as [Carpinone et al., 2010]
nij
pij = Pk
(2.62)
nijl
,
pijl = Pk
l=1 nijl
(2.63)
j=1 nij
and
respectively, where n is the number of occurrences for a given state.

If pij = 1, the state i is said to be an absorbing state, which means that
if the process enters this state, it cannot leave it [Balzter, 2000]. A state is
either [Ross, 2010]:
Recurrent: If the process is in state i, the probability of returning to
this state is 1,
Transient: If the process is in state i, the probability of returning to
this state is less than 1.
Further the Markov chain is said to be irreducible if it all states communicate, i. e. accessible to each other.
By the Chapman-Kolmogorov equations, the probability of going from
state i to state j in t steps, known as the t-step transition probability is
defined as [Ross, 2010]
ptij = P (Xt+k = j|Xk = i),
36
t, k 0,
i, j 0
(2.64)
2.9. MARKOV CHAINS

and
t+m
pij
=
otik om
kj ,
t, m 0,
i, j,
(2.65)
k=0
t+m
where pij
is the t + m step transition probability of going from state i to
state j. This can also be expressed using matrices so that
P (t+m) = P (t) P m ,
where P (t) is the t-step transition matrix.
37
(2.66)
CHAPTER 2. THEORY
38
Chapter 3
Methods
3.1
Site and time
Fakken wind farm is located at Vannya in Troms county in Northern Norway. A map of the area is given in Figure 3.1. The wind farm consist of 18
Vestas V90 3.0MW horizontal axis wind turbines [Troms Kraft AS, 2012]
with hub height of 80 meter and a rotor diameter of 90 meter with three
turbine blades [Vestas Wind Systems A/S, 2013]. All the turbines are sited
on the southwestern tip of Vannya, with a mountain range located to the
west, going across the island from south to north. Surrounding Vannya
are several other mountainous islands and fjords, resulting in a complex
wind regime. Both local and regional mountain waves are expected at the
site. The local mountain waves as a result of the local mountain regions on
the islands itself, and the regional as an effect of the great mountain range
between Norway and Sweden to the southwest.
Three sets of wind data have been provided and used in this study,
1. In-situ measurements from meteorological station located at Fakken
wind farm (courtesy of Troms Kraft AS),
2. In-situ measurements from a reference wind turbine, WTG08, at Fakken
wind farm (courtesy of Troms Kraft AS),
3. Numerical weather prediction data obtained by the AROME-Norway
model for the Fakken wind farm area (courtesy of the Norwegian Meteorological Institute),
with coordinates for the site, meteorology station and WTG08 given in Table
3.1. In-situ measurements are given for the time period 1. May 2013 to 1.
May 2014, while the AROME-Norway forecasts are given for the time period
between 2. May 2013 and 31. March 2014.
39
CHAPTER 3. METHODS
WTG08
Meteorological Station
Fakken wind farm
Vannya
Troms
Figure 3.1: Map of the location and surrounding area of Fakken wind farm
with WTG08 and meteorology pointed out.
Euroref 89
Fakken windfarm
Coordinates
UTM-zone 33
65660E
7828400N
710680E
7768400N
Met. station
Coordinates
Kote
Height [m]
693269.567E
68
80
7785030.852N
WTG08
Coordinates
Kote
Hub height [m]
693339.50E
66.13
80
7785249.78N
Table 3.1: Site coordinates
40
3.2. DATA COLLECTION
3.2
3.2.1
Data collection
In-situ measurements from VestasMetPanel 3000
The meteorological station VestasMetPanel 3000 measures the wind speed

and wind direction at 45 and 80 meter heights. From these measurements
the following data are available,
10-minute minimum value
10-minute average value
10-minute maximum value
10-minute variance
for both the speed and direction parameters [Vestas Wind Systems A/S,
2008]. Each record is time-stamped at the start of each record, i.e. data
recorded from 13:30:00 to 13:40:00 will be time-stamped 13:30:00, but stored
at 13:40:00.
Anemometer
Wind speed measurements at the meteorological station are performed using
two WAA151 anemometers, installed at the respective heights. An image of
these wind anemometers are seen in Figure 3.2 with the cup wheel assemble,
sensor shaft and lower body pointed out.
As described by Vaisala Oyj [2002a] the anemometer WAA151 is an optoelectronic, fast-response and low threshold anemometer. The wind speed
is measured by three light-weight conical cups which are attached to the cup
wheel. According to the producer, this will provide excellent linearity for
the entire operating range, which goes up to 75 m/s. Attached to the cup
wheels shaft is a chopper disc and as the wind rotates the cup wheel, this
chopper disc will cut an infrared light beam 14 times per revolution. This
generates a pulse train output from a phototransistor.
The output pulse rate makes it possible to determine the wind speed as
they can be regarded as directly proportional, for instance 246 Hz = 24.6
m/s. However, in order to increase the accuracy of the measurements the
characteristic transfer function is advised to be used as noted in Table 3.2.
To avoid problems related to freezing in cold climates, a heating element
of usually 10 W is installed in the shaft tunnel to keep the temperature of
the bearings above freezing levels. Key technical specifications are shown in
Table 3.2.
41
CHAPTER 3. METHODS
Figure 3.2: WAA151 anemometer, where (1) is the cup wheel assembly, (2)
the sensor shaft and (3) the lower body [Vaisala Oyj, 2002a].
Property
Measuring range
Starting threshold
Accuracy (within 0.4 to 60
m/s):
With Characteristic Transfer
function
With simple transfer function Uf = 0.1xR
Dimensions
Weight
Description/Value
0.4 - 75 m/s
< 0.5 m/s 1)
0.17 m/s
2)
0.5 m/s
240 (h) x 90 () mm
570g
1) Measured with cup wheel in position least favored by flow direction.
Optimum position yields < 0.35 m/s starting threshold.

2) Standard Deviation
Table 3.2: Key technical specifications for WAA151 Anemometer [Vaisala

Oyj, 2002a].
42

Wind Vane
In addition to the two anemometers, two WAV151 wind vanes are installed
at the same heights on the meteorology station. These are used to measure
the wind direction. The WAV151 wind vane is a counter balanced, low
threshold optoelectronic wind vane [Vaisala Oyj, 2002b]. Figure 3.3 shows
the WAV151 wind vane with the vane assembly, sensor shaft and lower body
pointed out.
Wind direction are measured using infrared LEDs and phototransistors
which are mounted on six orbits on each side of a 6-bit GRAY-coded disc
[Vaisala Oyj, 2002b]. As the wind direction changes, the vane is rotated
causing the disc to create changes in the code received by the phototransistors. In order to eliminate ambiguities, the code is changed in steps of 5.6 ,
one bit at a time.
A 10 W heating element is installed in the shaft tunnel to avoid the
temperatures of the bearings to reach freezing levels. Key technical specifications are shown in Table 3.3.
3.2.2
In-situ measurements from WTG08
For the WTG08 wind turbine, following measurements of the wind speed
and power have been provided
10-minute minimum value
10-minute average value
10-minute maximum value
10-minute variance
in addition to
10-minute average value of the temperature
10-minute absolute average value of the direction.
These date are collected by equivalent measurement instruments as described for the VestasMetPanel 3000.
3.2.3
Numerical weather prediction data
Hourly point prediction data are obtained from the numerical weather prediction (NWP) model AROME-Norway. Forecasts issued at 00 UTC-time
are provided, with a forecast horizon of 24 hours. These data includes prediction of the wind speed and wind direction of Fakken wind farm for two
layers, L64 and L65.
43
CHAPTER 3. METHODS
Figure 3.3: WAV151 wind vane, where (1) is the vane assembly, (2) the
sensor shaft and (3) the lower body [Vaisala Oyj, 2002b].
Property
Sensor/Transducer type
Measuring range
Starting threshold
Accuracy
Dimensions
Weight
Description/Value
Vane/Optical code disc
0 - 360
< 0.4 m/s
Better than 3
300(h) x 90() mm
Swept radius of vane: 172 mm
660g
Table 3.3: Key technical specifications for WAV151 Wind Vane [Vaisala Oyj,
2002b].
44

Height of each layer varies with the weather but in the general the L64
layer is approximately 34 meters above the ground, while the L65 layer is
approximately 10 meters above ground. For Fakken wind farm, the model
height is approximately 20 meters above sea level, but in reality the wind
farm lies 57 meter above sea level. The L64 and L65 layers are therefore in
practice at 54 and 30 meters above sea level respectively, and this must be
taken into account when extrapolating to hub height.
Rewriting the power law given by Equation 2.7 in Section 2.3.2, it is
possible to derive an expression for the Hellman constant a as follows
a
z
u(z) = u(zr )
zr
a
u(z)
z
=
u(zr )
zr

z
u(z)
= a ln
ln
u(zr )
zr ,
(3.1)

u(z)
ln u(z
r)

a =
ln zzr
a=
ln(u(z)) ln(u(zr ))
ln(z) ln(zr )
where a now can be estimated as the average of the measurements from the
two heights on the meteorological station or by using the predictions from
the L64 and L65 layers, expressed mathematically as

n
1 X ln(u(z80,i )) ln(u(z45,i ))
a
met.station =
,
(3.2)
n
ln(z80 ) ln(z45 )
i=1
and
a
f orecast
1X
=
n
i=1
ln(u(zL64,i )) ln(u(zL65,i ))
ln(zL64 ) ln(zL65 )

,
(3.3)
respectively. Here n is the total number of data points, i. e. measurements

or forecasts, and u(z80,i ) is the ith measurement of the wind speed at 80
meters height, equivalently u(zL64 ) is the ith wind speed forecast at 34
meters height.
Using the estimated Hellman exponent the two forecast layers L64 and
L65 can be extrapolated to hub height using the power law, with z = 117m
and zr = 34m for the L64 layer and z = 117m and zr = 10m for the L65
layer. The z height is calculated as follows
z =hW T G08 + hF akken hmodel
=80m + 57m 20m
=117m
45
(3.4)
CHAPTER 3. METHODS
where hW T G08 is the hub height of WTG08, hF akken the approixmately
height of Fakken wind farm above sea level and hmodel the height of the
AROME-Norway model for Fakken wind farm.
At the initiating of each forecast, the first hours will be influenced by
some noise and the longer forecast horizon the more uncertain the forecast
will be.
3.3
Power prediction models
3.3.1
Notation
The following notations will be used when dealing with the power prediction
models.
P : Measured power from WTG08 [kW]
P : Power forecast [kW]
U : Measured wind speed from meteorological station[m/s]
: Wind speed forecast using NWP [m/s]
U
D: Measured wind direction from meteorological station [deg]
Wind direction forecast using NWP [deg]
D:
t0 : Current time [hrs]
tm : m = -1, 1, 2, time at -1, 1 and 2 hours from current time.
3.3.2
Persistent model
The persistent model is often used as a reference model for evaluating shortterm power prediction models. It is given by Carpinone et al. [2010] as
P (th+m |th ) = P (th ),
(3.5)
for m = , , , ... and where P (th+m |th ) is the power forecast for time th+m
made at time th and P (th ) is the measured power at time th . So in words
this model simply states that future power production will equal the last
measured power. However, even though it is a simple predictor, it will
perform rather well for the first look-ahead times due to the atmospheric
changes occurring relatively slow [Madsen et al., 2005].
46
3.3. POWER PREDICTION MODELS
3.3.3
Training and test data
In order to find out how well a prediction model perform, it should be tested
on new and independent test data [Madsen et al., 2005]. The available data
should therefore be divided into two parts, a training period, and a test
period, of which the prediction model is trained using the training data
before the capability of the model is tested using new and independent data
from the test period.
This method will not only be a more correct way of finding the performance, but it will also easily show cases of over-training. If the models
complexity increases, it will become more capable of distinguishing between
the input data, leading to good performance when the model is tested on the
same data it was trained with. However, such a model will perform poorly
when tested on new independent data, as it is over-trained [Madsen et al.,
2005].
In this study, the leave one out method will be used to provide both
training data and new, independent test data. Each test period will consist
of a month of data, with the remaining 10 month being training data. Error
measures will only be based on the test period.
3.3.4
Markov chain model
An overview of the algorithm for the Markov chain model are shown in
Figure 3.4. All the 10-minute averaged data are loaded from xlsx-format into
Matlab with an array for each of the various input data, such as measured
wind speed, power output and so on.
In the next step the 10-minute average data are converted to hourly data.
The measured power output are converted as hourly arithmetic averages,
while for measured wind speed and direction the last 10-minute average of
every hour were chosen, as they are closest to the point predictions provided
by the NWP. This means that the data measured between 10:50 and 11:00,
timestamped as 10:50, was used as the hourly data value for 11:00. At the
same time, all the measured data and forecast data are made sure to have
the same relative time, which was chosen to be UTC-time.
Further all missing dates and NaN-values (no available measurements)
are removed from all the arrays so that each array consist of the same
amount of data for the same time. Negative values are set equal to 0 and
power values above rated power are set to rated power. This is mainly
done to simplify the definition of the states later on. All forecast data are
extrapolated to hub height using the power law and the estimated Hellman
constant a using the average between Equation 3.2 and Equation 3.3.
For each of the three input types, power P [kW], wind speed U [m/s]
and wind direction D [deg], number and spacing of state intervals are chosen.
These have currently been chosen to be
47
CHAPTER 3. METHODS
P: 20 states, of which 19 are equally spaced between 0 and 3000 kW,
with an additional cut-off states being defined when wind speeds are
above 10 m/s and power output equals zero,
U: 9 states, of which 0 - 3.5 m/s (cut-inn speed) is the first state, the
next 7 states are between 3.5, 5, 7.5, 10,..., 15, 20, 25 and the last
state being all wind speed above cut-off speed. Note that two states
are given between the rated wind speed at 15 m/s and the cut-off speed
at 25 m/s,
D: 8 equally spaced states representing the cardinal directions north,
east, south, west and the intercardinal directions northeast, southeast, southwest and northwest. Each direction covers 45 , with 22.5
on each side of the directional centers. (These directions can be seen
in Figure 4.3 in Chapter 4).
For every data array, each element of the vector is converted to its corresponding state value, for instance a wind speed value of 0 m/s would be
denoted as state 1, while a wind speed value of 30 m/s denoted as state 9
(as this is the total number of states for wind speeds). This is done due
to practical reasons in the algorithm, making it both more computational
effective and easier to handle several input parameters to the model later
on.
The arrays, now with each element representing the state according to
their corresponding parameter state interval, are then divided into test and
train periods. For each of the 11 months, 11 test arrays are made containing
data for a whole month, with corresponding 11 arrays consisting of data
from the remaining 10 months, to be used as a training data for making the
transition matrix.
For each Markov chain model, a set of parameters must be defined as
input to the model. This could for instance be current measured wind speed
(t2 )) or any other
(U (t0 ))and 2-hour point forecast of the wind speed (U
combination from the available data. It is possible to also define the output
state parameters, however this is set to be the power prediction with a 2-hour
prediction horizon by default, i. e. a forecast of the average power output
between the next hour ,t1 , and two hours into the future, t2 .
Now that all the prerequisite are taken care of, the actual power prediction can be made. Starting by the first test month, a transition matrix is
made using data from the remaining training data. With the example above,
using as input parameters the current measured wind speed and the 2-hour
point forecast and the predicted power output in 2-hours, the transition
matrix P would look become
48
3.3. POWER PREDICTION MODELS
p1,1,1
p1,2,1
..
.
p1,u,1
P =
p2,1,1
p2,2,1
.
..
p1,1,2
p1,2,2
..
.
p1,u,2
p2,1,2
p2,2,2
..
.
p1,1,p
p1,2,p
..
..
.
.
p1,u,p
,
p2,1,p
p2,2,p
..
..
.
.
(3.6)
pu,u,1 pu,u,2 pu,u,p

where pijl indicates the probability of going to the power output state l,
given that the current measured wind speed is in state i and the 2-hour
point forecast in state j. Using the default number of states defined above,
the dimension of the matrix would be 81x20, i. e. number of wind speed
states to the second power times the number of power output states.
For the test month, the current state of the first element of the input
parameter arrays are found. These values indicates a given row in the transition matrix, which can be used to find the probabilistic forecast of the
power output, i. e. the column elements of that particular row. Currently
the model give as output the most probable power state, which again can
be converted back to a specific power value by using the definition of the
power states. In the case that no previous observations have been recorded
for a specific event, i. e. the row in the transition matrix consists only of 0s,
0 power output will be used as the forecast.
This procedure is done for every hourly data point in the test month,
repeated for a new test month and new training months until all the data
have been used for testing and training. This results in a complete set of
independently tested data. The NRMSE are calculated both monthly and
for the whole period.
3.3.5
Combined Markov chain and persistent model
The combined Markov chain and persistent model use the MC model by
default, but in the case of no previously recorded data for a given row in the
transition matrix, the persistent model with m = 2 is used instead. This
means that instead of choosing the 0 power output prediction the average
power output for the last hour will be used instead.
49
CHAPTER 3. METHODS
Load data
from xlsx-file
Pre-process
data
Define states
Convert to
state vectors
Split data into
test and train
period
Define input
parameters
Predict
power output
Figure 3.4: Algorithmic overview of the Markov chain wind power prediction
model.
50
Chapter 4
Results
Using Equation 3.2 and Equation 3.3 the Hellman constant was estimated to
be a
met.station = 0.11 and a
f orecast = 0.09 and therefore a Hellman constant
of a = 0.10 was chosen for extrapolation of forecast data from the layer
height to hub height. Such low values of a implies a rather smooth surface.
In this case the solution of the power law (Equation 2.7) will go towards the
solution of the logarithmic law (Equation 2.3) [Emeis, 2013]. The power law
was therefore chosen for the extrapolation process, due to its mathematical
simplicity. As the extrapolation of the two layers L64 and L65 was found to
be relatively equal, only extrapolated values for the L64 layer will be used
in the proceedings. All in-situ measurements used are from 80 meter height.
As wind forecasts are only available in the time period 2. May 2013 to
31. March 2014, the same time period is used for the in-situ measurements
from Fakken in the following sections. This is to ensure the same amount of
data from the same time period, to be able to directly compare measured
and forecast data in the proceeding sections. An exception is Section 4.3
were only the in-situ measurements are used in the analysis. In this section
the time period 1. May 2013 to 31. April 2014 are used instead.
4.1
Rose diagrams
The wind distribution as a function of direction for measurements from the

meteorology station at Fakken are given in Figure 4.1, and for the weather
forecast in Figure 4.2.
Starting with Figure 4.1, the most common wind directions seen are between southeast and southwest and from the north-northwest, with southeast being the most dominant direction. Winds from the southeast have
a higher occurrence of high wind speeds compared to the other directions.
For the other directions the wind distribution between each sector are quite
similar.
Looking at the rose diagram of the wind forecast (Figure 4.2) some dras51
CHAPTER 4. RESULTS
tic changes can be observed. Still the most dominant wind direction is from
southeast, with most of the predicted winds from between a southeasterly
and a southwesterly direction. However the winds are now more evenly
spread within these sectors and not decreasing from southeast to southwest
as was the case of the real measurements. In addition the high wind speeds
seen from the southeast above 12.5 m/s are either not present or drastically
reduced between the two rose diagrams. Further the northwesterly winds
seen in Figure 4.1 are now seen as northern winds.
Figure 4.3 shows a map of the surrounding areas of Fakken wind farm
(red rectangle) with the wind directions displayed on a circle around the
map. Comparing with the wind directions seen in Figure 4.1 it is possible to observe that the dominant winds from the southeast corresponds
with the strait between Arnya and Lenangsyra. Further the southern
winds corresponds to the strait between Reinya and Lenangsyra, while
the southwestern directions points to the straits and area between Vannya,
Ringvassya and Reinya. Winds blowing from the north-northwest are
winds coming in from the ocean, following the curvature of the mountain
range on Vannya. Arnya covers almost the entire eastern direction while
the mountain range at Vannya covers both the western and part of the
southwestern and northwestern directions.
Comparing with Figure 4.2 the weather forecast predicts more winds
blowing in from the open ocean to the north, lower occurrence of winds
from the strait between Arnya and Lenangsyra and instead more winds
blowing from direction of Lenangsyra, Reinya and Ringvassya.
4.1.1
Case study
Figure 4.4 and Figure 4.5 shows a case study of 27. October 2013 for measurements from the meteorology station and weather forecasts respectively.
In the first figure the winds throughout the day are mainly blowing from
a southeastern direction with most wind intensities being in the range 5 - 15
m/s and some in the range 15-20 m/s. The second figure have much lower
wind intensities with most being in the range 0 - 12.5 m/s. Still the dominant
wind direction is to the southeast, but a higher number of predicted winds
are also seen to the east.
4.2
Power curve
Power curves for the WTG08 are shown in Figure 4.6 and Figure 4.7. The
former shows the power output of WTG08 as a function of measured wind
speeds (blue circles), while the latter as a function of forecast wind speed
(red circles). In both figures the theoretical power curve is given in black
with the cut-in, rated and cut-off wind speeds shown as lines at 3.5 m/s, 15
m/s and 25 m/s respectively.
52
4.2. POWER CURVE
NORTH
15%
12%
9%
6%
3%
WEST
EAST
Wind speed [m/s]

25 99
20 25
15 20
12.5 15
10 12.5
7.5 10
5 7.5
3.5 5
0 3.5
SOUTH
Figure 4.1: Rose diagram of in-situ measurements from the meteorology

station at Fakken wind farm, in the period 2. May 2013 - 31. March 2014
(UTC-time).
NORTH
15%
12%
9%
6%
3%
WEST
EAST
SOUTH
Wind speed [m/s]

25 99
20 25
15 20
12.5 15
10 12.5
7.5 10
5 7.5
3.5 5
0 3.5
Figure 4.2: Rose diagram of the forecast wind at Fakken wind farm, in the
period 2. May 2013 - 31. March 2014 (UTC-time).
53
CHAPTER 4. RESULTS
N
NE
NW
Vannya
Arnya
Ringvassya
Reinya
SW
Lenangsyra
SE
S
Figure 4.3: Fakken wind farm (red rectangle) with the surrounding area and
directions shown.
54
4.2. POWER CURVE
NORTH
15%
60%
12%
40%
9%
6%
20%
3%
WEST
EAST
Wind speed [m/s]

25 99
20 25
15 20
12.5 15
10 12.5
7.5 10
5 7.5
3.5 5
0 3.5
SOUTH
Figure 4.4: Rose diagram of in-situ measurements from the meteorology

station at Fakken wind farm, 27. October 2013 (UTC-time).
NORTH
15%
60%
12%
40%
9%
6%
20%
3%
WEST
EAST
SOUTH
Wind speed [m/s]

25 99
20 25
15 20
12.5 15
10 12.5
7.5 10
5 7.5
3.5 5
0 3.5
Figure 4.5: Rose diagram of the forecast wind at Fakken wind farm ,27.
October 2013 (UTC-time).
55
CHAPTER 4. RESULTS
In Figure 4.6 the measured values follows quite well the same trend as
the theoretical power curve, with some distinguished outliers observed. The
majority of measurements are seen below 1000 kW and 10 m/s.
For the power curve using forecast wind speed the same trend can be
seen, but is observed to be drastically shifted towards lower wind speeds.
This is especially seen by the high number of rated power observations occurring at very low wind speeds. This corresponds with the lower wind speeds
that was seen in the rose diagrams between measured and forecast data. In
addition a higher spread is observed over the whole specter, especially below
1000 kW and 10 m/s compared to Figure 4.6.
4.2.1
Case study
A case study of 27. October 2013 are shown in Figure 4.8 and Figure 4.9,
using the measured and forecast wind speed respectively. In these plots it
is much easier to see the difference between real and predicted data. For
measured values, the data points lies close to the theoretical power curve
with a relatively small spread, while for the forecast values, all the data
points are located to the left side of the theoretical curve. However, the
forecast data still follows the same trend. It should also be noted that for
this particular day almost the entire specter of wind speed and power can
be observed.
4.3
Power, speed and direction diagrams
This section will consider how the power output from the WTG08 wind
turbine changes with respect to variations in the speed and direction of the
wind. All the data discussed are converted from their original 10-minute
values to hourly values by applying formulas corresponding to the statistical
features used. Data are recorded for one whole year, starting at 1. May 2013
and ending on the 31. April 2014.
In the following figures, there are images with equally sized rectangles
which illustrates a given direction and speed interval. Each of these boxes
corresponds to a given state or event and the color represent the intensity
measured. Dark red means higher intensity, while dark blue means a low
intensity. If a rectangle is white, no observations of this particular event
were recorded.
Be aware that the first and the three last wind speed states on the x-axis
have a broader speed interval, thus a higher number of recordings from these
states could be expected. The reason for splitting these states up in this
order is to be able to easier analyze special features of the power production
occurring at cut-in, cut-off and rated wind speeds. It should also be noted
that all data corresponding to 0 directional measurements were omitted.
This is because in meteorology, 0 means that there is not high enough
56
4.3. POWER, SPEED AND DIRECTION DIAGRAMS
3500
3000
Power [kW]
2500
2000
1500
1000
500
0
0
10
15
Wind speed [m/s]
20
25
30
Figure 4.6: Theoretical power curve (black) with measured power output
from WTG08 as a function of measured wind speed (blue circles). Time
period is 2. May 2013 - 31. March 2014 (UTC-time).
3500
3000
Power [kW]
2500
2000
1500
1000
500
0
0
10
15
Wind speed [m/s]
20
25
30
Figure 4.7: Theoretical power curve (black) with measured power output
from WTG08 as a function of forecast wind speed (red circles) .Time period
is 2. May 2013 - 31. March 2014 (UTC-time).
57
CHAPTER 4. RESULTS
3500
3000
Power [kW]
2500
2000
1500
1000
500
0
0
10
15
Wind speed [m/s]
20
25
30
Figure 4.8: Case study of 27. Ocotber 2013 (UTC-time). Theoretical power
curve (black) with measured power output from WTG08 as a function of
measured wind speed (blue circles).
3500
3000
Power [kW]
2500
2000
1500
1000
500
0
0
10
15
Wind speed [m/s]
20
25
30
Figure 4.9: Case study of 27. Ocotber 2013 (UTC-time). Theoretical power
curve (black) with measured power output from WTG08 as a function of
forecast wind speed (red circles) .
58
4.3. POWER, SPEED AND DIRECTION DIAGRAMS
>100
N
80
70
60
SW
50
S
40
SE
30
20
Number of Measurements
Wind Direction
90
NW
10
NE
0
0 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 20 25 25<
Wind Speed [m/s]
Figure 4.10: The number of measurements observed for a given wind speed
and direction in the period 1. May 2013 - 31. April 2014 (UTC-time), using
hourly averaged data.
wind speed for a wind direction to be measured, and should therefore not
be considered in this case. When it is possible to measure a direction, 360
is used instead.
4.3.1
Number of observations
First consider Figure 4.10, which shows the number of observations recorded
in the time period for a given wind speed (x-axis) and wind direction (y-axis)
state. This is similar to the rose diagrams seen in Section 4.1, however it is
for a broader time period, as the data in this section is not considering the
weather forecasts, and the data points are now obtained as averages over an
hour.
The distribution of the observations show some clear trends. In general
low wind speeds occur in all directions with a high number of observations
available. The definitely most dominant wind directions are from south
and southeast followed by northwest and southwest. Winds from eastern
and especially northeastern directions are not occurring frequently and high
speed winds from northeast are not recorded in the time period at all. Low
wind speeds are observed more frequently than high wind speeds.
4.3.2
Standard deviation
The standard deviation (STD) of the measured power output are shown in
Figure 4.11. This is a statistical measure for how much the observations
deviate from each other, i. e. how much spread there is in the recorded data.
A higher value means a greater spread.
59
CHAPTER 4. RESULTS
300
250
W
200
SW
150
S
Power [kW]
Wind Direction
NW
100
SE
E
50
NE
0
0 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 20 25 25<
Wind Speed [m/s]
Figure 4.11: The standard deviation of the observed power output for a
given wind speed and direction in the period 1. May 2013 - 31. April 2014
(UTC-time), using hourly averaged data.
The most interesting feature in this plot is the high STD seen for westerly winds and to a high degree also for winds blowing from southwest and
northwest. A high variation are also observed for wind speeds in the states
around 25 m/s for several of the wind directions. In general a higher STD
are seen for higher wind speeds, decreasing with decreased speed. Power
output for the four bottom rows, directions going from south to northeast,
are more or less stable over the entire wind speed range.
4.3.3
Averages
Two methods for analyzing the average of the power output as a function
of wind speed and wind direction are seen in Figure 4.12. In the upper
subplot the arithmetic mean has been used to calculate the average, while
the median is used in the lower subplot. The reason for choosing both
of these averaging methods are to better illustrate the average of the power
output. For normally distributed data the arithmetic mean is a good method
for finding the average of the sample, however this method is very sensitive
to outliers in the sample, i. e. skewed distributions. In this case the median
will give a better estimate of the average. Note that in these subfigures
two additional rows are added to the y-axis. One is intentionally left empty
while the other shows the power curve for the WTG08 wind turbine, which
can be used as a reference when analyzing the data.
As expected, the general trend is that the power output increases for
an increased wind speed until the wind reaches the cut-off speed, where
the power production suddenly drops. The power output in the lower wind
speed specter follows the power curve quite well for all directions. At about
60
4.4. MARKOV MODEL

7 - 10 m/s some outliers are seen, which actually have a higher power output
than what should be expected from the power curve.
For wind speeds above this however, more variations in the power output
can be observed. Here the power output seems to generally be lower than
the corresponding values seen for the power curve. The row which deviates
the most from the power curve is the row corresponding to western winds.
In this direction the rated power output are not seen in either the mean or
the median plot for any wind speeds and the power output is relatively far
below the expected output.
At the interval between 14 - 14.5 m/s for a western direction, only two
observations have been recorded and both at zero values. This suggest that
a special event have occurred while these measurements were recorded.
Comparing the two columns representing the rated power output, i. e.
wind speeds of 15 - 25 m/s, for both the mean and median, the first column
corresponds well with the power curve, while the second column have a
noteworthy lower average power output. An exception is in the case of
easterly and southeasterly winds, where only the mean values are lower,
while the median values are at the rated power. These two cases will be
studied in more detail in the following case study.
4.3.4
Case study
In this case study two states will be looked into for a more detailed analysis.
The first is the western wind direction for the wind speed interval 15 - 20
m/s and the second state is for the southestern direction, with the same
wind speed interval.
Figure 4.13 shows the distribution of the observed wind power, using the
arithmetic mean for averaging.
For the upper plot, i. e. southesterly winds, the power output is observed
within the range of 2900 - 3000 kW in almost 70 % of the cases, with the
rest of the observations being close to the rated output. In the lower plot
however, the distribution is completely different. Here 15 % of the observed
winds have zero power, with rated power output recorded in less than 15
% of the total observations. Most of the wind power measurements are
seen above 1750 kW, but much more evenly spread out compared to the
southeastern direction.
4.4
4.4.1
Markov model
Leave one out method
A total of 18 Markov chain (MC) models with various input parameters

where tested using the leave one out method (see Section 3.3.3). Additionally
two persistent models (PM) were run for each month separately and for the
61
CHAPTER 4. RESULTS
3000
N
NW
2500
2000
SW
S
1500
SE
E
Power [kW]
Wind Direction
1000
NE
500
Power Curve
0 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 20 25 25<
Wind Speed [m/s]
(a) Mean.
3000
N
NW
2500
2000
SW
S
1500
SE
E
Power [kW]
Wind Direction
1000
NE
500
Power Curve
0 3.5 4 4.5 5 5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.5 11 11.5 12 12.5 13 13.5 14 14.5 15 20 25 25<
Wind Speed [m/s]
(b) Median.
Figure 4.12: The mean (a) and median (b) power output observed for a given
wind speed and direction in the period 1. May 2013 - 31. April 2014 (UTCtime), using hourly averaged data. Last row displays the corresponding
power curve for the WTG08 turbine.
62
4.4. MARKOV MODEL
70
Occurrence [%]
60
50
40
30
20
10
0
0
500
1000
1500
Power [kW]
2000
2500
3000
2000
2500
3000
(a) Southeast
25
Occurrence [%]
20
15
10
0
0
500
1000
1500
Power [kW]
(b) West
Figure 4.13: The distribution of mean wind power observed for wind speeds
in the interval between 15 - 20 m/s for winds from southeast (a) and west
(b) in the period 1. May 2013 - 31. April 2014 (UTC-time), using hourly
averaged data.
63
CHAPTER 4. RESULTS
whole period, to be used as reference models. The first PM use m = 1 in
Equation 3.5, i. e. a model which states that the average power output over
the next hour equals the average power output over the last hour. Second
PM use m = 2, which means that the average power output over the last
hour equals the power forecast of between next hour and two hours into the
future. The latter is the model currently used to forecast power production
at Fakken wind farm and is therefore of special interest for comparison. By
definition, no training period are required for the PM (see Section 3.3.2).
Combining each independently tested month, it was possible to obtain a
dataset containing test data for the whole period. For every model, the normalized root mean square error (NRMSE) were calculated for the combined
dataset and the results are seen in Table 4.1.
As expected, M1 performs better than M2, however as the power production companies must give the expected power output between t1 and t2
at time t0 , only M2 will be applicable in practice and this is the reason why
the current model used for Fakken wind farm is M2. Even so, it is still
interesting to use M1 as a comparison.
M3 to M11 are MC models with various input parameters, while M12
to M20 have the same input parameters, but in this case the MC models
are combined with M2. These combined models use the MC as default, but
if the current state used to predict P (t ) are not found in the transition
matrix, i. e. no such event have been previously recorded, the model will
instead use P (t ) as a forecast of P (t ), which is the M2 persistent model.
The default option for MC would be to choose state 1 instead, which is a
zero power output state.
Of the MC models, M4 and M7 have the lowest NRMSE over the entire
time period and their equivalents M13 and M16 have the lowest NRMSE
of the combined models. These models outperform the currently used M2,
only by using measured and forecast wind speed to predict the future power
output. The best performing model, M16, has a 2.24 % lower NRMSE value
than M2 during the time period tested and has only got a 3.46 % higher
NRMSE value than M1.
The worst performing models are M10 and M11, which consist of many
input parameters of measured and forecast wind speed and direction, with
NRMSE values of 25.70% and 27.54% respectively.
In general higher NRMSE can be observed for the MC models compared
to their equivalent combined models, with exception of M4 and M13.
Figure 4.14 shows the NRMSE as a function of months for the M1, M2,
M4 and M7 models. Comparing the models with each other it can be noted
that M1 beat the three other models in performance for every month, as
would be expected from the table values seen in Table 4.1, but with M4
and M7 being really close in July and August. The persistent models M1
and M2 are observed to have relatively equal NRMSE curves throughout
the time period. In general the M4 and M7 performs better than the M2
64
MC + M2
MC
Type
PM
65
12
13
14
15
16
17
18
19
20
1
Pt1 , Pt0
Ut1 , Ut0
Ut1 , Ut0 , Pt0
t , U
t , Pt
U
0
2
0
t
Ut1 , Ut0 , U
2
Ut1 , Dt1 , Ut0 , Dt0
t ,D
t
Ut0 , Dt0 , U
2
2
t ,D
t
Ut1 , Ut0 , Dt0 , U
2
2
t ,D
t
Ut ,Dt ,Ut , Dt , U
Pt1 , Pt0
Ut1 , Ut0
Ut1 , Ut0 , Pt0
t , U
t , Pt
U
0
2
0
t
Ut1 , Ut0 , U
2
Ut1 , Dt1 , Ut0 , Dt0
t ,D
t
Ut0 , Dt0 , U
2
2
t ,D
t
Ut1 , Ut0 , Dt0 , U
2
2
t ,D
t
Ut1 ,Dt1 ,U,t0 , Dt0 , U
2
2
Pt0
Pt0
1
2
3
4
5
6
7
8
9
10
11
Input Parameters
Model #
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt2
Pt1
Pt2
Output Parameters
22.49
17.23
17.57
20.96
16.84
18.16
17.79
18.89
19.11
23.07
17.20
18.75
22.58
17.58
20.52
21.02
25.70
27.54
13.38
19.08
NRMSE (%)
4.4. MARKOV MODEL
Table 4.1: Overview of the performance of persistent models , Markov chain

models and combined Markov chain and persistent models (MC + M2) for
various input parameters.
CHAPTER 4. RESULTS
35
M1 (13.38%)
M2 (19.08%)
M4 (17.20%)
M7 (17.58%)
NRMSE (% of Nominal Power)
30
25
20
15
10
0
May13 Jun13
Jul13
Aug13 Sep13 Oct13 Nov13 Dec13 Jan14 Feb14 Mar14

Month & Year
Figure 4.14: The NRMSE of M1, M2, M4 and M7 shown for each month
with the NRMSE for the whole time period given in parenthesis behind the
respective model legend.
model, especially in July and August 2013 and February and March 2014.
From September to January M2, M4 and M7 have relatively equal NRMSE
values.
Looking at how the performance changes with time, all models are seen
to perform better in the late spring and summer months compared to the
autumn and winter months. An exception is January, which breaks the
general trend.
Corresponding curves for the combined models are seen in Figure 4.15,
with M13 and M16 plotted with the reference models M1 and M2. The
trends seen here are similar to those seen in Figure 4.14 with only some
minor changes. M16 perform better in November and December than was
the case for the equivalent M7 model and this is also the case for February
and March, where M16 now are relatively equal to M13.
4.4.2
Case study
A case study of the power forecasts of model M2, M4 and M16 compared to
the real power output for 27. October 2013 (UTC-time) are given in Figure
4.16. Model M7 and M13 were emitted in this case, as their plots were not
significantly different from their equivalent models M16 and M4 for this day.
Starting with M2, it is easy to see that this curve is a 2 hour delayed
version of the real power output. For small changes the M2 model forecast
66
4.4. MARKOV MODEL
35
M1 (13.38%)
M2 (19.08%)
M13 (17.23%)
M16 (16.84%)
NRMSE (% of Nominal Power)
30
25
20
15
10
0
May13 Jun13
Jul13
Aug13 Sep13 Oct13 Nov13 Dec13 Jan14 Feb14 Mar14

Month & Year
Figure 4.15: The NRMSE of M1, M2, M13 and M16 shown for each month
with the NRMSE for the whole time period given in parenthesis behind the
respective model legend.
relatively good the real power output, for instance seen in the time periods
01 to 05, 11 and 12 and 15 to 19. However, for greater changes and variation
in power output, the M2-model deviates more from the real curve. This is
seen especially when the power output changes relatively fast from rated
output to low power values in only a few hours around 06 to 09. In general
the M2 are observed to predict lower values than what was measured at 13
and 14 and at 20 to 24, with an exception at 22.
The M4 model dwells longer on the rated power output, compared to
M2 and M16, before it suddenly decrease drastically the forecast in just a
few steps. It underestimate the power output at similar times as the M2,
namely at 13, 14 and 20 to 24, also with an exception at 22 where it predicts
accurately the power output.. Also at 15 to 17 the M4-model overestimate.
The M16 model manage to follow the rapid decrease in power output
during the nighttime and morning to a higher degree than what was seen
for M2 and M4. It should also be noted that it lies at a lower power from
start, compared to the other models. Like the previous models, M16 also
underestimates the power output at 13 to 14, but at 15 to 17 it overestimate quite much more than the M4-model, It underestimate at 19 and 20,
overestimates at 22, before going back to underestimating the power output
at 23 and 24.
67
CHAPTER 4. RESULTS
3000
Real
M2 (19.08%)
M4 (17.20%)
M16 (16.84%)
2500
Power [kW]
2000
1500
1000
500
0
01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
UTCtime 27. October 2013
Figure 4.16: Case study of 27. October 2013 (UTC-time) showing the predicted power output from the M2 (black), M4 (red) and M16 (green) models
compared to the measured power output (blue). The NRMSE corresponding
to each model for the whole time period are given in parenthesis.
68
Chapter 5
Discussion
5.1
Wind roses
In general the results from Section 4.1 show that the wind seem to favor
directions corresponding to straits between the many surrounding islands
of Vannya. A reason for this are likely due to the high mountains and
mountain ranges located on the islands, making the straits less resistant
pathways for the wind.
For the measured data, the rose diagram showed a good portion of the
wind blowing from north-northeast. This suggests that the winds blowing
in from the ocean follows the curvature of the mountain range on Vannya,
crossing the island from the north-northwest to the south.
Few winds were observed from east and west compared to the other
directions. This is likely the result of Arnya to the east working like an
obstacle to any natural wind flows from this direction, and the mountain
range to the west of Fakken.
The results also showed that the predicted and measured wind direction
had some major differences. This could be due to the resolution of the NWP
from AROME-Norway. Even though the resolution is greatly improved from
previous models, a resolution of 2.5 km will still have some difficulties in accurately recreating the topography of areas with huge variations and rapid
changes in height, such as mountain ranges. The topography in the model
would therefore be smoothed, reducing the obstacle-effects which could be
seen from mountain ranges. For instance the AROME-Norway model predicts Fakken wind farm at 20 meter above sea level, while the true height
is approximately 57 meter. This could also be the reason for the shifting of
direction from northeast-north to north from the measured to forecast wind
direction.
69
CHAPTER 5. DISCUSSION
5.2
Power curve
As expected the power curve for the measured wind speed data does not
follow the theoretical power curve exactly, but the trend is still clear. Main
reason for this being that the meteorological station does not measure perfectly the wind speed incident on the wind turbine. Many factors could
possibly contribute to measurement errors. Take for instance the variation
in the wind direction, in some cases the meteorological station will be ahead
of the turbine making good measurements of the incident wind, while in
other cases, the direction of the wind could cause the meteorological station
to be at the lee side of the turbine, which would result in the wind already
having a reduced energy. The meteorological station would then be shadowed by the WTG08 turbine, as was discussed for wake effects in Section
2.6.1. Other factors could be local turbulence, wake effects, wind turbine
settings and so on.
It should also be noted that both measured and forecast data are for 80
meter height, while the rotor of the wind turbine has a diameter of 90 meter,
making the turbine blades span from 35 to 125 meter heights. A power curve
should therefore not be used as a direct link between wind speed and the
power output, but more of a indicator of the expected power output for a
given speed.
The power curve from forecast wind speed gives a very clear picture of
the general underestimation of wind speeds from the NWP model, this is
especially seen for all the rated wind power measurements occurring at low
wind speeds. This is also seen for the case study. However, it does follow
the same trend, though with a larger spread. It could therefore be possible
to use a correction factor to try and reduce the effect of underestimation.
5.3
Power, speed and direction diagrams
Like with the rose diagrams seen in Section 4.1, Figure 4.10 also show that
the wind patterns corresponds well with the observed straits between the
islands surrounding Fakken seen in Figure 4.3.
The standard deviation values seen in Figure 4.11 showed some interesting results. Winds blowing from westerly direction had significantly higher
STD-values compared to other directions, suggesting a higher degree of unstable winds. This could very well be related to the mountain wave phenomena discussed in Section 2.6.2, which would mean that winds passing
over the mountain range to the west would be more turbulent and therefore
create a higher spread in the measured power output.
Even so it should also be noted that the states with a higher number of
observations does seem to have a lower STD, vice versa, comparing Figure
4.10 and Figure 4.11. A contradiction to this hypothesis is the eastern and
70
5.4. MARKOV CHAIN MODELS

northeastern direction where both low number of measurements and low
STD-values are observed. However, as low number of observations may
easily skew the STD-measure, it would be wise to look at some more data
before drawing any conclusions regarding the mountain waves.
For the last two wind speed states in the STD-figure high values are
observed for most of the directions. This could be expected, as these states
corresponds to the cut-off speed where small changes in wind speed may
lead to a sudden drop in power output due to the turbine being shut-down
to avoid fatigue and for safety reasons.
In Figure 4.12 the mean and median of the power output were shown.
Rated power output were not observed at all for winds from the west in
any of the subfigures, and neither from northwest or southwest in the mean
subfigure. This means that there is definitely some incidents occurring for
westerly winds. Also in this case the observation fits very well with the
mountain wave phenomena, as more turbulent winds would lead to difficulties in obtaining efficient power production and generally give a lower power
output, as seen here.
Variations between the mean and median values was especially seen for
wind speeds from 15 - 25 m/s. These variations can be divided into two
cases. The first case is when the median shows a rated power output, while
the mean shows a lower power output. In this case there are a few outliers
in the data sample which pulls down the mean estimate, while the majority
of the measurements are observed with the rated power output. This will
cause the mean to be reduced, while the median will still show the rated
power output. In the second case, reduction in both the mean and median
values are observed. For this to happen, the majority of the observations
must have a lower power output than the rated power, as only this may
cause the median to drop.
This corresponds very well with what was seen in Figure 4.13. These
distributions clearly show that for southeastern wind directions the power
output are much more consistent, with the majority of values being close to
the rated power. The few outliers will naturally pull down the mean value,
while the median will indicate rated output. For western winds however,
the distribution is much more spread with many occurrences of what could
be cut-off wind speeds.
The case study also clearly shows the occurrence of turbulent winds from
a western direction and how this affects the power production.
5.4
Markov chain models
In general the MC models have higher observed NRMSE values than the
combined MC and M2 models. This indicates that there are occurrences
where the currently observed state have not been recorded before, leading
71
CHAPTER 5. DISCUSSION
to the MC model prediction zero power output, while MC + M2 model use
the current power output. Generally lower NRMSE for the combined models
suggest that the current power output is typically closer to the real power
output in 2 hours than the zero power output state. Exception are the M4
and M13 models. In this case the guess of zero power output is most likely
closer to the real power than the current power production.
Another trend seen in the Table 4.1 are how the NRMSE depends on
the number of input parameters, which is equivalent to the size of transition
matrix. M10 and M11 have the highest NRMSE and also the highest number
of input parameters and combined states. As discussed in Section 3.3.3 this
probably cause the model to be over-trained. In general too many input
parameters and number of states for these parameters would increase the
complexity of the model, making it less capable of predicting new values,
while too few input parameters or too few states for these would lead to
under-training of the model and also poor performance.
Take for instance M11, with three wind speed and three wind direction
input parameters. Combined with the power output parameter, these would
make a 373248X20 transition matrix, with 7 464 960 possible states! Of
course in practice, all these states would not be real possible events, making
some of them redundant. However, it would require vast amounts of data
the fill all the probable states. Even for a small transition matrix, such is
the case for M4, it would be hard to fill all the elements of the matrix with
observations using only measurements from 11 months. See for instance
Figure 4.10 where no high speed winds from northeast have been observed,
even though they are possible states in the transition matrix. From this it
is possible to conclude that a Markov chain model would be more accurate
for a higher number of observations used to fill the transition matrix.
In Figure 4.14 and Figure 4.15 the M2 model is observed to follow the
same trends as the M1 mode. Reasons for this is probably due to both using
the same predictor, namely the persistent model.
An interesting observation is seen between the M7 and M16 models in
the two figures. For November, December, February and March, the M16
perform much much better than the M7, indicating that in these months
a lot of observation which has not been recorded before in the transition
matrix are seen. Looking at the NRMSE curves from a seasonal perspective,
it seems to be generally lower prediction errors for months characterized by
more calm and stable weather, such as late spring and summer months, while
more errors are made when the autumn and winter weather steps in, usually
consisting of much more unstable weather. January 2014 was observed as a
extraordinary dry and calm month, with almost no precipitation, which may
be the reason why all models made relatively good wind power predictions
for this month. This seasonal trends may also be the cause of the M16
performing better than M7, as unstable weather could very well indicate
weather conditions which have not previously been recorded.
72
Chapter 6
Conclusions
A statistical analysis was conducted for Fakken wind farm based on measurements recorded in the time period 1. May 2013 - 31. April 2014 and NWP
data in the time period 2. May 2013 - 31. March 2014. This showed that the
main wind directions are from the straits between Arnya and Lenangsyra
to the southeast and between Rainya and Lenangsyra to the south. Winds
blowing in from the ocean to the north and northeast are observed to follow
the mountain range on Vannya, leading to north-northwesterly winds at
Fakken wind farm. Few winds were observed from the west and east. This
indicates that Arnya to the east and the mountain range on Vannya, west
of Fakken, are working as obstacles for the wind flows from these directions.
Only a few observations were made of winds from northeast, mainly being
low speed winds below 10 m/s.
Wind forecast showed similar trends, but the wind speed were generally underestimated and the wind directions were more evenly distributed
and skewed. This can be concluded as being a result of how the AROMENorway NWP model the topography, reducing the affects from the surrounding mountains.
High standard deviation of the power output from the WTG08 reference
turbine was observed for winds mainly from west, but also from northwest
and northeast. In addition low mean and median values were observed for
westerly winds, indicating that winds from the west are more turbulent than
any other direction. This is most likely the result of mountain waves occurring for winds crossing over the mountain range on Vannya, significantly
influencing the power production.
A total of 18 power prediction models were developed, of which 9 were
based on Markov chains and the remaining 9 were combined models based
on Markov chains and a persistent model. The performance were evaluated using the normalized root mean square error and compared with two
persistent models.
The currently used power prediction model at Fakken wind farm, a per73
CHAPTER 6. CONCLUSIONS
sistent model with a prediction horizon of m = 2, was found to have a
NRMSE of 19.08%. Several of the models developed in this study were
found to have lower NRMSE than the currently used model, of which the
best performing model had a NRMSE of 16.84 %. This model use as input
the currently measured wind speed and in the previous hour, together with
forecast wind speed in 2-hours. This 2.5% lower NRMSE corresponds to approximately 3.1 106 kWh of the anually electricity production from Fakken
wind farm.
These discoveries indicates a great potential in using power prediction
models based on Markov chains going directly from various input parameters, such as measured and forecast wind speeds, to predicted power output.
6.1
Further research
There are several features which could improve the performance of the wind
power prediction models discussed in this study.
Currently the states are more or less linearly spaced for each parameter
and does not take into consideration how the available observations are
distributed. It could therefore be beneficial to look into various clustering
methods and how they could be used to create more ideal states.
A high number of input parameters and/or a high number of state intervals results in a high number of empty states with no observations recorded.
This will cause problems when evaluating new data, as these might not have
been previously observed, leading to the MC models choosing a 0 kW wind
power forecast by default. It is suggested to replace this default behavior with a method for finding the nearest neighbor in the transition matrix
instead.
It would also be very interesting to combine the MC models with the
work of Yoder et al. [2013], using the suggested model of 1 hour ahead
categorical change in wind power as an indicator for the MC models.
Only one of the four daily weather forecast issued by AROME-Norway
have been used in this study. The first hours of every forecast are usually
characterized by a higher degree of noise. Combining forecasts issued at
different times, would make it possible to avoid this noise and obtain more
accurate wind forecast.
74
Bibliography
John Andrews and Nick Jelley. Energy Science, principles, technologies, and
impacts. Oxford University Press, 2007.
Trygve Aspelien and Morten Kltzow. AROME-Norway From experiments to official public forecasts for the whole wide world. [Online],
April 2013. URL http://www.cnrm.meteo.fr/aladin/IMG/pdf/ASM_
AROME-Norway201304_TA.pdf. Accessed on May 16th 2014 at 10:12.
Holger Babinsky. How do wings work? Physics Education, 38(6):497503,
2003.
Heiko Balzter. Markov chain models for vegetation dynamics. Ecological
Modelling, 126(2):139154, 2000.
Arthur Bossavy, Robin Girard, and George Kariniotakis. Forecasting ramps
of wind power production with numerical weather prediction ensembles.
Wind Energy, 16(1):5163, 2013.
Godfrey Boyle. Renewable Energy. Oxford University Press, 2004.
BP. BP statistical review of world energy june 2013, 2013.
A. Carpinone, R. Langella, A. Testa, and M. Giorgio. Very short-term
probabilistic wind power forecasting based on Markov chain models. pages
107112, 2010.
K. Conradsen, L. B. Nielsen, and L. P. Prahm. Review of Weibull statistics
for estimation of wind speed distributions. Journal of Climate and Applied
Meteorology, 23(8):11731183, 1984.
Philippe Drobinski and Corentin Coulais. Is the weibull distribution really
suited for wind statistics modeling and wind power evaluation? arXiv
preprint arXiv:1211.3853, 2012.
Stefan Emeis. Wind Energy Meteorology Atmospheric Physics for Wind
Power Generation. Springer-Verlag Berlin Heidelberg, 2013.
EREC. Renewable energy the solution to climate change, 2005.
75
BIBLIOGRAPHY
Douglas C. Giancoli. Physics principles with applications. Pearson Education International, sixth edition, 2005.
Gregor Giebel, Richard Brownsword, George Kariniotakis, Michael Denhard, and Caroline Draxl. The state-of-the-art in short-term prediction
of wind power: A literature overview. Technical report, ANEMOS. plus,
2011.
F. Gonz
alez-Longatt, P. Wall, and V. Terzija. Wake effect in wind farm
performance: Steady-state and dynamic behavior. Renewable Energy, 39
(1):329338, 2012.
Zack Guido. Climate dynamics. [Online], September 2008. URL http:
//www.southwestclimatechange.org/climate/global/dynamics. Accessed on 11.05.2014 at 15:05.
Robert V. Hogg and Allen T. Craig. Introduction to Mathematical Statistics.
Macmillan Publishin CO.,Inc., fourth edition edition, 1978.
James R. Holton and Gregory J. Hakim. An introduction to dynamic meteorology. Academic press, 2012.
John Houghton. Global Warming the complete briefing. Cambridge University Press, fourth edition edition, 2009.
IPCC. Summary for policymakers. in: Climate change 2014: Impacts, adaption, and vulnerability. Contribution of Working Group II to the Fifth
Assessment Report of the Intergovermental Panel on Climate Change,
2014.
Morten Jacobsen. Statistical analysis of wind velocities at Fakken wind farm
on Vannya in Troms., 2013. Project paper, UiT The Arctic University
of Troms.
S. Rao Jammalamadaka and Ambar Sengupta. Topics in circular statistics,
volume 5. World Scientific, 2001.
C. G. Justus, W. R. Hargraves, and Ali Yalcin. Nationwide assessment
of potential output from wind-powered generators. Journal of Applied
Meteorology, 15(7):673678, 1976.
F. Koch, M. Gresch, F. Shewarega, I. Erlich, and U. Bachmann. Consideration of wind farm wake effect in power system dynamic simulation. In
Power Tech, 2005 IEEE Russia, pages 17. IEEE, 2005.
Alberto Leon-Garcia. Probability and Random Processes for Electrical Engineering. Addison-Wesley Publishin Company, Inc, second edition edition,
1994.
76
BIBLIOGRAPHY
Henrik Madsen, Pierre Pinson, George Kariniotakis, Henrik Aa Nielsen, and
Torben S. Nielsen. Standardizing the performance evaluation of shortterm wind power prediction models. Wind Engineering, 29(6):475489,
2005.
Robert P. Mahan. Circular statistical methods: Applications on spatial and
temporal performance anaylsis. Technical report, United States Army
Research Institute for the Behavioral and Social Sciances, April 1991.
Knati V. Mardia and Peter E. Jupp. Directional Statistics. John Wiley &
Sons Limited, 2000.
Andrei Andreyevich Markov. Extension of the limit theorems of probability
theory of a sum of variables connected in a chain. The Notes of the imperial
Academy of Sciences of St. Petersburg, VIII Series, Physio-Mathematical
College XXII, (9), 1907.
Melinda Marquis, Jim Wilczak, Mark Ahlstrom, Justin Sharp, Andrew
Stern, J. Charles Smith, and Stan Calvert. Forecasting the wind to reach
significant penetration levels of wind energy. Bulletin of the American
Meteorological Society, 92(9), 2011.
Sathyajith Mathew. Wind Energy Fundamentals, Resource Analysis and
Economics. Springer-Verlag Berlin Heidelberg, 2006.
Dr. Brad Muller. Mountain and lee waves in satellite imagery. [Online], May 2014. URL http://wx.db.erau.edu/faculty/mullerb/
Wx365/Mountain_waves/mountain_waves.html. Accessed on 15.05.2014
at 18:46.
Norwegian Meteorological Institute. Nettet sner seg sammen. [Online], February 2008. URL http://met.no/?module=Articles;action=
Article.publicShow;ID=727. Acessed on May 16th 2014 at 10:33.
Norwegian Meteorological Institute. N
a blir vrvarslene enda bedre.
[Online], October 2013. URL http://met.no/N%C3%A5+blir+v%C3%
A6rvarslene+enda+bedre.b7C_xdfY4e.ips. Accessed on May 16th 2014
at 10:17.
Ca A. Paulson. The mathematical representation of wind speed and temperature profiles in the unstable atmospheric surface layer. Journal of
Applied Meteorology, 9(6):857861, 1970.
S. A. Pourmousavi Kani and M. M. Ardehali. Very short-term wind speed
prediction: A new artificial neural networkMarkov chain model. Energy
Conversion and Management, 52(1):738745, 2011.
77
BIBLIOGRAPHY
Hans-Holger Rogner, Roberto F. Aguilera, Cristina L. Archer, Ruggero
Bertani, S.C. Bhattacharya, Maurice B. Dusseault, Luc Gagnon, Helmut
Haber, Monique Hoogwijk, Arthur Johnson, Mathis L. Rogner, Horst
Wagner, Vladimir Yakushev, Doug J. Arent, Ian Bryden, Fridolin Krausmann, Peter Odell, Christoph Schillings, and Ali Shafiei. Energy resources
and potentials.
Sheldon M. Ross. Introduction to Probability Models. Elsevier Inc., 10th
edition, 2010.
Peter J. Schubel and Richard J. Crossley. Wind turbine blade design. Energies, 5:34253449, 2012.
J. V. Seguro and T. W. Lambert. Modern estimation of the parameters of
the weibull wind speed distribution for wind energy analysis. Journal of
Wind Engineering and Industrial Aerodynamics, 85(1):7584, 2000.
Y. Seity, P. Brousseau, S. Malardel, G. Hello, P. Beenard, F. Bouttier,
C. Lac, and V. Masson. The arome-france convective-scale operational
model. Monthly Weather Review, 139(3), 2011.
A. Shamshad, M. A. Bawadi, W. M. A. Wan Hussin, T. A. Majid, and
S. A. M. Sanusi. First and second order markov chain models for synthetic
generation of wind speed time series. Energy, 30(5):693708, 2005.
SSB. Minifacts about norway 2013. [Online], May 5th 2014. URL http://
www.ssb.no/a/english/minifakta/en/main_11.html. Acessed on May
16th 2014 at 10:35.
Henry Stark and John W. Woods. Probability, Statistics and Random Processes for Engineers. Pearson Education Limited, fourth edition edition,
2012.
Troms Kraft AS. Fakken vindpark. [online], July 2012. URL http://www.
tromskraft.no/om/prosjekter/fakken. Acessed on June 2nd 2014 at
14:52.
Vaisala Oyj. USERS GUIDE Anemometer WAA151. P.O Box 26, FIN00421 Helsinki Finland, 2002a. URL http://www.vaisala.com.
Vaisala Oyj. USERS GUIDE Wind Vane WAV151. P.O Box 26, FIN-00421
Helsinki Finland, 2002b. URL http://www.vaisala.com.
Vestas Wind Systems A/S. VestasMetPanel 3000 Meteorological Station
General Specification. Alsvej 21 8900 Randers Denmark, April 2008. URL
http://www.vestas.com.
Vestas Wind Systems A/S. V90 3.0MW. Hedeager 44, 8200 Aarhus N,
Danmark, 2013. URL http:\www.vestas.com.
78
BIBLIOGRAPHY
Ronald E. Walpole, Raymond H. Myers, Sharon L. Myers, and Keying Ye.
PROBABILITY & STATISTISCS for ENGINEERS & SCIENTISTS.
Pearson Educational International, eight edition edition, 2007.
Xiaochen Wang, Peng Guo, and Xiaobin Huang. A review of wind power
forecasting models. Energy procedia, 12:770778, 2011.
Megan Yoder, Amanda S. Hering, William C. Navidi, and Kristin Larson.
Short-term forecasting of categorical changes in wind power with markov
chain models. Wind Energy, 2013.
79

Markov Chain and Numerical Weather Prediction Models

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Markov Chain and Numerical Weather Prediction Models

Загружено:

Авторское право:

Доступные форматы

Department of Physics and Technology

Short-term wind power prediction based on Markov

Atmospheric boundary layer

HAWT Horizontal axis wind turbine

NRMSE Normalized root mean square error

RMSE Root mean square error

Rotations per minute

A simple convection cell, with low pressure denoted by L and

Rose diagram of in-situ measurements from the meteorology

Time scale of forecast models and their applications [Wang

Overview of the performance of persistent models , Markov

Short-term wind power prediction

The stochastic nature of wind makes it difficult to develop models which

1.2. FORMER RESEARCH

- Economic load dispatch planning

Multiple days ahead

Forecast models used to predict more specific events or scenarios of power

1.3. PURPOSE OF THE STUDY

Purpose of the study

Structure of the study

Source of wind energy

Wind energy, or simply wind, is the kinetic energy of air in motion. As

Global wind patterns

2.2. GLOBAL WIND PATTERNS

Figure 2.2: The global wind circulation [Guido, 2008].

Vertical wind speed profile

2.3. VERTICAL WIND SPEED PROFILE

However, rarely is the thermal stratification of the Prandtl layer found

where L is the Obukhov length defined as

2.4. DISTRIBUTION OF WIND

for 0 < z/L 0.5

B (z/L C/D) exp (D(z/L ))

with a = 5, A = 1, B = 2/3, C = 5 and D = 0.35 [Emeis, 2013].

Figure 2.4: Weibull probability density function as a function of wind speed

The maximum likelihood method

2.4. DISTRIBUTION OF WIND

where M = u2 + v 2 is the wind velocity with wind speed components u

R the weight 2corresponding to channeled flow events occurrence and Z =

Wind power production

Wind energy is the kinetic energy of air in motion. As described by Boyle

2.5. WIND POWER PRODUCTION

Starting with the definition of power it is possible to derive an expression

Second expression is obtained by looking at the change of momentum.

Combining Equation (2.21) and Equation (2.23), an expression for u

2.5. WIND POWER PRODUCTION

Using Equations (2.24) and (2.25), Equation (2.23) becomes

For a = 1, P = 0, and the maximum power is therefore found for the

Figure 2.5: Airflow through a HAWT.

Wind turbine blade design

2.5. WIND POWER PRODUCTION

Figure 2.6: Air particles flowing along an airfoil [Babinsky, 2003].

(a) Low angle of attack

(b) High angle of attack

(c) Stalled flow

an increased speed of the fluid, vice versa [Giancoli, 2005].

2.5. WIND POWER PRODUCTION

Horizontal axis wind turbine design

(b) Turbine blade section.

2.5. WIND POWER PRODUCTION

Figure 2.9: Design of a modern HAWT [Schubel and Crossley, 2012].

Cutout wind speed

Cutin wind speed

2.6. TOPOGRAPHIC FEATURES

Mountain waves are a feature of mesoscale meteorology which occur when